malaiwah/quant-fidelity-registry

Quantization Fidelity Registry

A public, schema'd, receipt-backed, cross-model index of quantization quality measurements.

It exists to answer one question that nothing else answers today: show me every measured quant of model X, with its fidelity number and enough provenance to know whether that number means anything.

It is the sibling of 0xSero/local-ai-registry, which answers how fast, how much VRAM, how much money. This one answers how faithful. Ids and the huggingface identity block are deliberately shaped so records can be cross-linked; every artifact, model, panel and pipeline carries a cross_refs.local_ai_registry slot, and a link is never presented as verified unless it has been.

Community test fixtures

The registry includes twelve native random fixture bases with qualified public CPU root datasets and exact-zero two-capture reproduction controls. Twenty linked storage-format artifacts retain their actual shared-weight parents and public capture/comparison receipts. These are toolchain fixtures, not trained assistants or model-quality rankings; floating-point conversion controls are not fine-tunes.

• Random Architecture Fixtures
• Matched-Weight Quantization Families
• Native root captures and exact measured source

The fixture registry rows are advisory relative to the production sealed GPU lane: their CPU reproduction is exact, but no cross-lane offset is invented. Metadata and source licenses remain distinct; bundled source license texts retain their own terms.

publication-audit.json retains every validator warning with an explicit disposition. This snapshot has no schema/invariant errors, but it is not a warning-free software release: historical source-version drift, missing controls/secondary metrics and other stated limitations are preserved. The strict check-release gate is unchanged; no warning is silently made green.

---

The rule this registry exists to enforce

Two fidelity numbers with different comparability.key values are NEVER comparable. Equal keys make two rows candidates for comparison — a necessary condition, not a certificate. Before ranking rows as like-for-like, check the group's machine-readable comparability predicate in index.json.

(Until 2026-08-31 this rule was stated as an "if and only if". The only-if half was always true and still is. The if half was never true, and an independent peer review said so with this registry's own receipts: the key deliberately omits the measurement lane — one artifact in group cmp--202b717f3219c414 is measured on two lanes, differing in the fourth decimal; the candidate pipeline — the same checkpoint/panel/teacher has measured ~24% apart through two pipelines; hardware — a measured A100-vs-H200 term of 2.97e-4 nats is ~13x the gap between two published 4-bit quantizers (ARCHITECTURE-DETERMINISM); and artifact scope — a routed-experts-only quant and a full-forward GGUF at "the same" bpw are different interventions. The omission is deliberate — hashing hardware into the key was considered and rejected, because it would explode the partition into single-row groups — but deliberate is not the same as sufficient, and the contract now says which it is. The key itself is unchanged and unversioned: rehashing it would regroup every published row and orphan every key a third party has cited. What changed is the claim made about it, plus a machine-readable predicate per group: comparable: true/false/unknown with reasons, in index.json, recomputed by the validator (CMP-007) so a hand-edited predicate is rejected exactly like a forged key.)

The secondary check also reads declared replay backend/environment and stack evidence. Missing evidence stays unknown; a backend or stack mismatch is not erased by a shared pipeline name. A metric-covering harness match certifies the recorded code closure; different harness IDs alone do not prove different numerics and remain unknown.

A bare kld: 0.027 is worse than nothing. A KL divergence is only meaningful relative to a specific set of tokens, measured against a specific teacher capture, in a specific direction, at a specific accumulator precision, through a specific stack relation, with a specific head policy. Change any one of those and you have a different quantity that happens to be printed with the same units.

So the key is a hash over exactly those seven things:

comparability.key = "cmp--" + sha256("|".join([
    panel_id,            # WHICH TOKENS, including the scored-position policy
    reference_id,        # WHICH TEACHER CAPTURE (a capture, not a model: artifact + panel + stack + precision)
    metric_name,         # mean_tokenwise_kld, mean_of_run_means_tokenwise_kld, ...
    direction,           # reference_to_candidate (KL(P_teacher || Q_student)) or the reverse
    accumulation_dtype,  # float64 vs float32 over 10M positions is a different estimator
    stack_relation,      # same_stack or cross_stack; the latter's bias direction may be unknown
    head_policy,         # the candidate's own lm_head, or one shared head applied to both sides
]))[:16]

tools/registry_validate.py recomputes this key for every row from the row's own fields and rejects a mismatch (CMP-001). A hand-written key cannot move a number into a table where it does not belong. tools/registry_render.py groups tables by that key and by nothing else, and --check fails if the committed README differs from what the data renders. The tables below are a pure function of data/*.jsonl. They were never typed by hand and cannot drift.

A worked example: descriptive values and an invalid cross-key ranking

Five numbers, all for GLM-5.3-Flash, all on brandonmusic's sealed 25-window / 51,175-position panel, all against the same stored fp32 teacher logits, all KL(teacher || student) in nats, all accumulated in float64. They are printed as two tables, not one, because they are two quantities. A reader who skims tables rather than paragraphs should be stopped by the layout, not only by the prose underneath it:

Group cmp--202b717f3219c414 -- sealed-lane same-stack capture, five cold runs each. These are historical panel values, not a certificate of like-for-like ranking: apply the pair predicate, including pipeline and provenance evidence, before ranking them.

	value	metric	stack_relation
malaiwah TR3 6bpw (K6), 253.5 GB	0.013723384665701147	mean_of_run_means_tokenwise_kld	same_stack
brandonmusic tr3 4bpw, 175.6 GB	0.024554564249958208	mean_of_run_means_tokenwise_kld	same_stack
0xSero EXL3 Q4 (Dione), 187.6 GB	0.027262784814670614	mean_of_run_means_tokenwise_kld	same_stack

Group cmp--4a8630bdcadab97f -- a different quantity, not a continuation of the table above. Single-pass cross-stack replay against that same stored teacher. The unquantized control contextualizes the FP8 value; it is not a competing quantization result.

	value	metric	stack_relation
BF16 replay (the floor)	0.012711599817250710	mean_tokenwise_kld	cross_stack
official FP8 (our replay)	0.020615254540417995	mean_tokenwise_kld	cross_stack

Note the sizes in the first table. K6 has the smallest reported value and is also the largest artifact in it by 66 GB. The 4bpw pair reports 0.024555 and 0.027263 at 175.6 GB and 187.6 GB. These are descriptive observations on the recorded panel; equal keys or similar nominal bit rates alone do not certify a fidelity ranking.

DESCRIPTIVE: "On this recorded panel the K6 row reports 0.013723, the brandonmusic 4bpw row 0.024555, and the Dione Q4 row 0.027263." The teacher and token receipts match. That establishes shared inputs, not shared candidate pipelines, replay arithmetic or runtime evidence. The pair predicate is required before turning these values into a like-for-like quantizer ranking.

INVALID: "The official FP8 release (0.020615) beats his 4bpw (0.024555) and loses to our K6." Different key -- and it differs on two axes at once. The FP8 number came from replaying the model through our vLLM stack and scoring it against a teacher captured on his transformers/eager stack. That is a cross_stack measurement, conflating runtime and quantization perturbations. Replaying the reference's own unquantized BF16 weights through our stack scores 0.012712 against those same teacher logits. This control does not determine the sign of the stack contribution to the FP8 value: KL is not additive and perturbations can cancel. 0.020615 is a descriptive cross-stack result, not an upper bound on quantization-only error. The scalar difference 0.007904 does not isolate that error. The control is labelled separately, never used as a mathematical lower bound or a quantization correction.

And one comparison the key alone does not stop

A comparability key has seven inputs and none of them is the measurement lane. Two rows can therefore share a key -- same panel, same teacher, same metric, same direction, same float64, same same_stack, same native_head -- and still have been produced on different machines by different code paths. Group cmp--202b717f3219c414 now contains exactly that: our K6 measured on the sealed 8x H200 lane at 0.013723384665701147, and the same weights measured on a one-GPU streaming lane at 0.013714888822596553. Sorted into one list, the streaming row lands above the sealed one and reads as a better quant. It is not a quant at all: it is one artifact, measured twice.

So the renderer tables a non-sealed lane's rows apart from the rest of its group, and the lane's pipeline record carries the measured bridge to the sealed lane rather than an adjective: signed delta -8.4958e-06 nats on the panel mean, worst single window 2.8735e-04, tokenwise_kld_sha256_matches_sealed: false, publishable_as_reproduction: false. The last two are the load-bearing ones. A mean that agrees to five decimal places is not a reproduction when the per-token array underneath it differs, and the lane says so about itself.

That bridge is one artifact's, on one panel, and it is not subtractable. The 8bpw row in the same lane has no sealed-lane sibling to bridge against, so its bias block records the offset as direction: unknown and its magnitude as null -- which is what "we do not know" looks like when it has to survive a schema.

The streaming lane also carries its own floor: the reference's own unquantized BF16 weights, scored through this SAME streaming harness rather than the cross-stack replay pipeline. It reads 0.011506 nats -- the cost of comparing across capture stacks plus bf16 non-associativity, with zero quantization involved -- and it is emphatically NOT the cross-stack floor above (0.012712, a different pipeline, a different lane, a different comparability key). The historical Excess over control (nats) column (formerly Attributable) is a descriptive scalar difference: K6-stream gives 0.002209 and K8-stream 0.000878. It does not identify a causal quantization effect, even when both inputs share a lane. No ratio of those residuals is warranted by these scalars; the old "2.52x" headline is withdrawn. BIAS-006 keeps floor references on the same lane, but passing that identity guard does not make KL additive. A value below an unquantized control is flagged for inspection, not rejected solely on that basis: cancellation can produce it legitimately. See engines/BF16-FLOOR.md for the full analysis.

The second differing axis is the metric itself: the K6 / 4bpw / Dione rows are mean_of_run_means_tokenwise_kld over five cold runs, while the cross-stack rows are a single mean_tokenwise_kld pass. When a measurement is bitwise reproducible those two coincide numerically, but they are not the same estimator in general -- brandonmusic's own v44 FP8 runs span 0.024016 to 0.024883 -- so the registry keeps them apart rather than assuming determinism it has not evidenced.

Ask the tool rather than reasoning it out yourself:

$ python3 tools/registry_validate.py \
    --explain measurement--glm53.k6-6bpw.brandonmusic-final25 \
    --against measurement--glm53.official-fp8.brandonmusic-final25.crossstack

NOT COMPARABLE. Differing comparability-key fields:
  metric_name         mean_of_run_means_tokenwise_kld
                      mean_tokenwise_kld
  stack_relation      same_stack
                      cross_stack
Everything else matches (panel_id, reference_id, direction, accumulation_dtype, head_policy).
measurement--glm53.official-fp8.brandonmusic-final25.crossstack declares bias.direction=upward with floor
measurement--glm53.bf16-replay-floor.brandonmusic-final25 (value 0.01271159981725071). Subtracting floors
is NOT sanctioned by this registry: the floor is context, not a correction.

That historical row's upward declaration is retained as historical metadata, not a theorem. New cross-stack submissions may honestly declare direction: unknown with usable_as_floor: false; downstream floor use is then explicitly refused.

A third case worth stating outright, because it is the one most likely to mislead: the MLX builds are measured against the official FP8 release dequantized to BF16, not against a BF16 teacher. Their 6-bit reads 0.0063, which is numerically smaller than our K6's 0.013723. It is not better. It is a different quantity -- the reference itself is quantized, so the FP8 error sits in the teacher instead of in the student. Those rows carry reference_kind: dequantized_from_quant, a mandatory different_reference_kind disclosure, and a panel marked undisclosed. They will never appear in a table with a native_bf16 row. No systematic ordering relative to a native-BF16 teacher follows: a quantized proxy can make KL either smaller or larger, depending on the candidate and reference.

---

What a row must carry

Every measurement names, and cannot validate without: a model and a pinned artifact; a panel (its own first-class record: corpus lineage, context and position counts, tokenizer, contamination guard, scored-position policy, token digest, availability); a reference -- modelled as a capture (artifact, panel, stack, logits precision, head source), so naming a teacher has already named a panel; a pipeline; the KL direction; the estimator precision; the run count with typed determinism evidence; the measurement scope; the provenance; the derived comparability key; and a non-empty disclosures array.

Three of those deserve emphasis, because they are where fidelity registries usually go wrong.

Determinism needs evidence, not a boolean. A receipt file's own sha256 proves nothing about numerics -- report files embed timestamps, paths and run indices, and differ across bit-identical runs. Only tensor content digests can back a determinism claim, and the schema blocks the rest (DET-001). The registry's own data is what taught it: in the K6 five-run receipt the five runs carry five different student_backend_identity_sha256 values (five genuinely distinct cold executions) and one identical tokenwise_kld_sha256. Container hashes would have said "nondeterministic"; tensor content says "bitwise identical". Conversely brandonmusic's v44 FP8 rows report five distinct tokenwise digests and a non-zero spread, and are recorded as not reproducible -- while his v44/v71/v75 NVFP4 rows report a single digest across five runs and are recorded as bitwise identical. Same author, same panel, opposite verdicts, both evidenced.

Who measured it is four separate facts, not one. provenance.measured_by is self-measured | author-reported | third-party-reported. independently_verified is a separate boolean that is never implied by it, and setting it true requires a verifier who is a different party than the measurer (PROV-003). Whether the artifact is ours is a third axis, carried by the third_party_artifact_self_measured disclosure -- the Dione Q4 row is 0xSero's weights and our number, and the table says exactly that. Whether the panel is ours is the fourth. Third-party numbers are welcome here and are never silently merged with ours.

Which code produced the number is a field, not a footnote. Every row carries a harness block: content digests of the computational closure that computed the value -- the estimator, its numerical support, the surface it read -- enumerated by role, plus the tool versions, reduced to one harness_id. Equal id means byte-identical code; a differing id points at the code_digests entry whose role changed. The commit sha is recorded beside it and deliberately excluded from the id, because a commit changes on a docs edit and an identity that churns for reasons that cannot move a number stops carrying information. The 72 rows that predate the mechanism (2026-08-30) are listed in schema/harness-grandfather.json, each carrying a harness_unrecorded disclosure; that list is frozen, and a new row without a harness is refused (HARN-001). They are not retroactively invalidated -- their receipts are hashed and their values reproduce -- and their digests are not reconstructed from a later checkout either, because today's files are not the files that produced them.

An assertion is a published claim, exactly as much as a number is. A metric row has always needed a hashed receipt. A provenance assertion -- "this bf16 twin is a direct cast, not a dequantization", "this NVFP4 config block is inherited from the parent rather than authored" -- needed nothing, and two such claims reached published dataset cards and registry rows with no source at all. A disclosure that makes one now sets asserts_provenance and carries its own pinned sources with an optional lines anchor (PROV-014/PROV-015/PROV-016). Pinned means a commit sha or a digest: /blob/main/ is refused outright, because line numbers move and a citation that quietly stops pointing at what it claimed still reads as evidence.

Panels are identified by their tokens, and the scoring window is part of that identity. Our GLM suite scored from position 0 gives 0.028104; the same tokens, the same artifact, the same teacher, scored from position 1024, gives 0.018794. A 33% move with nothing changed but which positions were averaged. So the second one is a separate panel record with derivation.kind: scoring_window_change, therefore a separate comparability key, therefore structurally unable to share a table with the first.

---

Provenance notes on this seeding

Two things in this data are worth stating plainly rather than burying in a disclosure.

brandonmusic's 25-window panel is genuinely sealed, and we verified it ourselves. Its identity is panel.json from his public teacher-logits dataset at revision 95f4fdd9, sha256 6bafe3283c54bc9342d0f30aa3199d36032d103feb92c31715be8545362790ff -- a manifest of 665 windows, each with its own token_ids_sha256, of which 25 carry role: final. That digest was recomputed by downloading the file during seeding and it matches the token_panel_artifact_sha256 his own panel receipt declares. The receipt's self-declared digest 0beec577... is recorded separately in identity.panel_receipt_sha256 and is explicitly barred from being used as a token identity or as determinism evidence (PANEL-002).

That panel's contamination guard is weaker than ours, and the tables say so. Its only guard is role separation: the 25 final windows come from the same packed corpus as the 384 fit / 128 conditional-fit / 64 selection / 64 confirmation windows, and are declared qualification-only. No lexical or n-gram scan is published. Our v5 suites run a 12-word shingle whole-document pre-exclusion against the calibration corpus and report 0 hits out of 941 documents scanned, 44 excluded. Those are not the same guard, and the validator warns whenever a strict row rests on a panel whose contamination.checked is false (PANEL-006). It applies equally to every row on that panel, so it does not disturb comparisons within it.

---

How to read the tables below

36 tables follow, one per comparability group, across 18 models. Three things are true of all of them, and each is a mistake somebody has already made with numbers like these:

1. A number means nothing outside its own table. Every table states the seven-part key its rows share. Two numbers under different keys are different quantities that happen to print in the same units.
2. The smallest number on this page is not the best quant. Today it is deepseek-v4-tiny-random-bf16 (random test fixture) native BF16 at 0 nats -- and it is not a quant at all -- those are unquantized weights, read by a second engine, measuring what two engines disagree by. Sorting this file by value and reading off the top is the single easiest way to be wrong with it.
3. Nothing here compares two models. A KL divergence is measured over one model's own vocabulary against that model's own teacher. GLM-5.3-Flash's numbers and Qwen3.8-27B's numbers are not on a shared scale and never will be.

Attribution is a column, not a footnote: measured by us, measured by us (their artifact) and reported by are three different epistemic states and the tables never merge them.

deepseek-v4-tiny-random-bf16 (random test fixture)

model--malaiwah.deepseek-v4-tiny-random-bf16 -- published by malaiwah. Tokenizer fixture-tokenizer-a411cee2863e41c1, vocabulary 272.

Panel: Synthetic CPU fixture panel 525cb6c62509

Panel disclosure -- reduced_run_count: Receipt disclosure reduced_run_count: this dataset is ONE cold capture (run_count 1, the DET-D4 floor is 5). Cross-run determinism is not asserted by the dataset; it is established by a second cold capture plus fidelity-dataset compare --self-compare, whose exactly-0.0 result is the SC-1 reproduction confirmation.

Panel disclosure -- native_head_replay: Receipt disclosure native_head_replay: HEAD-1d: each side replayed through its own sealed head (reference 5c5d4b4405a5, candidate 5c5d4b4405a5); head error is inside the measurement, as under HEAD-2, and nothing is substituted. The heads are content-identical.

Group cmp--b9834d82706b4192 -- 1 row

Panel panel--fixture.53326a9cd31fede1ea932442 -- Synthetic CPU fixture panel 525cb6c62509 4 contexts x 63 scored positions = 252 scored positions, score_from 0 sealed: yes (token digest 525cb6c625096ddb...) -- contamination scan: NOT RUN Reference (teacher) reference--fixture.deepseek-v4.bb5f2cdf5ea90c38 -- native_bf16, artifact artifact--malaiwah.deepseek-v4-tiny-random-bf16.185723bec10e @185723bec10ee57db86e8c269b3c6a811122a0ca Metric mean_tokenwise_kld, direction reference_to_candidate, accumulation float64 Estimation surface stack_relation same_stack, head_policy native_head Comparability key cmp--b9834d82706b4192 Like-for-like predicate comparable: unknown -- no recorded difference, but hardware, replay_backend, replay_env, stack are unrecorded for at least one member, so homogeneity cannot be certified. Machine-readable form with per-dimension values: this key's comparability block in index.json.

What this table is. Every row here shares the comparability key above: the same tokens, the same teacher capture, the same metric and direction, the same estimator precision, the same stack relation and the same head policy. That makes them CANDIDATES for ranking -- the key is a necessary partition, not a certificate. Whether they are also like-for-like on the dimensions the key omits (lane, pipeline, scope coverage, hardware) is what the predicate line above answers.

Rank is not a verdict. The table is sorted by fidelity alone, and fidelity buys bits: a larger, higher-bitrate quant will usually sit above a smaller one, which is not news. Read the Size and Codec columns before reading the order, and compare like against like.

What it is NOT comparable to. Every other table in this file: no other group shares this key. That includes every table for a different model -- a KL number is a divergence over one model's own vocabulary against that model's own teacher, never a score that can be carried between models.

Single-row group. This number has nothing in the registry to be ranked against. It is a stated fact, not a placing.

1 of this group's 1 rows came off a different measurement lane (other) and are tabled on their own below, not mixed into the ordering here. The key does not carry the lane; this file does.

Lane other -- 1 of this group's 1 rows

A different lane. Same key, and that is exactly the problem this table solves. The comparability key is a function of the panel, the teacher, the metric, the direction, the estimator precision, the stack relation and the head policy -- and these rows match the table above on all seven. What they do not share is the machine and the code path that produced the candidate logits, and lanes are not interchangeable. Sorting them into one list would read as a ranking; where the same artifact appears in both, it is one set of weights measured twice, not two quants.

No bridge to the sealed lane is recorded for this lane.

Artifact	Codec	Size	mean_tokenwise_kld (nats)	CI95	Top-1	Runs	Attribution	Receipt
deepseek-v4-tiny-random-bf16 (random test fixture) native BF16 (measurement floor)	bf16	0.0 GB	0	--	100.00 %	2 runs, bitwise identical	measured by us	receipt

Bias on deepseek-v4-tiny-random-bf16 (random test fixture) native BF16 -- other, direction unknown. This is the fixture's local CPU reproduction floor. Its offset against a production GPU lane was not measured; no cross-lane equivalence is claimed.

Disclosures for the rows above (3)

• fixture.deepseek-v4.floor.55d284c17c025a99 reduced_run_count: Receipt disclosure reduced_run_count: this dataset is ONE cold capture (run_count 1, the DET-D4 floor is 5). Cross-run determinism is not asserted by the dataset; it is established by a second cold capture plus fidelity-dataset compare --self-compare, whose exactly-0.0 result is the SC-1 reproduction confirmation.
• fixture.deepseek-v4.floor.55d284c17c025a99 non_sealed_lane: Local CPU fixture lane; this exact same-lane reproduction is not a measured offset against a production sealed GPU lane.
• fixture.deepseek-v4.floor.55d284c17c025a99 reduced_run_count: Two qualified cold runs, not five.

GLM-5.2-SIQ-Fruit

model--malaiwah.glm-5.2-siq-fruit -- published by malaiwah. Tokenizer glm-5.2-siq-fruit, vocabulary 154880.

Panel: Fruit held-out fidelity panel v1 -- 16 windows x 2048

Panel disclosure -- weak_contamination_guard: Separation from Fruit's training data is asserted at SOURCE level only: the two strata used are not among the nine sources Fruit's card names. No shingle or n-gram scan against the published pretraining shards was run, so incidental overlap through a web-crawl source such as FineWeb-Edu is not excluded.

Panel disclosure -- small_panel: 16 windows / 32,752 scored positions. On the one artifact measured here so far the per-window standard deviation is 0.0283 nats around a mean of 0.0387, a standard error near 0.0071. Numbers on this panel cannot separate artifacts that differ by less than roughly 30 percent.

Group cmp--e21ff3b61b1bb2ec -- 2 rows

Panel panel--fruit.malaiwah.heldout-v1 -- Fruit held-out fidelity panel v1 -- 16 windows x 2048 16 contexts x 2047 scored positions = 32,752 scored positions, score_from 0 sealed: yes (token digest a6d367cc3ba44880...) -- contamination scan: NOT RUN Reference (teacher) reference--malaiwah.fruit-bf16-hf.heldout-v1 -- native_bf16, artifact artifact--malaiwah.glm-5.2-siq-fruit-bf16 @ef68013aa6e16453cf52b5b77647f72fbe258c3c Metric mean_tokenwise_kld, direction reference_to_candidate, accumulation float64 Estimation surface stack_relation same_stack, head_policy shared_reference_head Comparability key cmp--e21ff3b61b1bb2ec Like-for-like predicate comparable: unknown -- no recorded difference, but harness, replay_backend, replay_env, stack are unrecorded for at least one member, so homogeneity cannot be certified. Machine-readable form with per-dimension values: this key's comparability block in index.json.

What this table is. Every row here shares the comparability key above: the same tokens, the same teacher capture, the same metric and direction, the same estimator precision, the same stack relation and the same head policy. That makes them CANDIDATES for ranking -- the key is a necessary partition, not a certificate. Whether they are also like-for-like on the dimensions the key omits (lane, pipeline, scope coverage, hardware) is what the predicate line above answers.

Rank is not a verdict. The table is sorted by fidelity alone, and fidelity buys bits: a larger, higher-bitrate quant will usually sit above a smaller one, which is not news. Read the Size and Codec columns before reading the order, and compare like against like.

What it is NOT comparable to. Every other table in this file: no other group shares this key. That includes every table for a different model -- a KL number is a divergence over one model's own vocabulary against that model's own teacher, never a score that can be carried between models.

Artifact	Codec	Size	mean_tokenwise_kld (nats)	CI95	Top-1	Runs	Attribution	Receipt
GLM-5.2-SIQ-Fruit BF16 (the reference export) (measurement floor)	bf16	10.1 GB	0	--	100.00 %	2 runs, bitwise identical	measured by us	receipt
GLM-5.2-SIQ-Fruit (exl3-trellis K3/K4 routed experts)	exl3-trellis @3.375	3.1 GB	0.0387375	--	87.98 %	1 run, unevidenced	measured by us	receipt

Disclosures for the rows above (5)

• fruit.siq-exl3-k3k4.heldout-v1 lossy_capture_codec: RECONSTRUCTED, NOT EXECUTED. The artifact's routed experts are exl3-trellis atoms that stock transformers cannot read, so the candidate capture ran a bf16 reconstruction of them (k6/tools/materialize_exl3_experts.py) rather than the vendor kernel. This is the dequantize-and-run methodology the GGUF/MLX/EXL3 ecosystems use for KLD: it measures the error of the STORED WEIGHTS and isolates it from kernel error. It does not measure Fruit's production path (b12x/SparkInfer + vLLM, fp8/nvfp4 KV, MTP). Decode evidence: the codebook table is bitwise equal to the campaign's independently frozen mcg table on all 65,536 entries; the bit rate read off every one of 8,448 payloads agrees with the producer's tier_bitmap; and the reconstruction error reproduces the ENCODER's own recorded expert_rel_rt_mse with ratio mean 1.00013 over range 0.98902-1.01337. The decode has NOT been proven bitwise against a running exllamav3 kernel, which is why this row is advisory.
• fruit.siq-exl3-k3k4.heldout-v1 small_panel: Per-window standard deviation 0.0283 around a mean of 0.0387 over 16 windows, i.e. a standard error near 0.0071. Do not rank this against anything it differs from by less than roughly 30 percent. The two strata differ by nearly 2x on their own (literary 0.0275, scientific 0.0500).
• fruit.siq-exl3-k3k4.heldout-v1 single_run: One cold capture of the candidate. Repeatability was established for the reference side only.
• fruit.siq-exl3-k3k4.heldout-v1 declared_scheme_mismatch: The artifact's config.json declares NVFP4/modelopt; the stored bytes are exl3-trellis K3/K4. scope_digest describes the bytes.
• fruit.siq-exl3-k3k4.heldout-v1 note: Per-window mean 0.038737453713514176, population sd 0.028308679654341876, min 0.012369540015856577 (final-0006, literary), max 0.09151472952402755 (final-0009, scientific) over 16 windows. The macro mean over contexts equals the token mean because every window contributes the same 2,047 positions.

glm-moe-dsa-tiny-random-bf16 (random test fixture)

model--malaiwah.glm-moe-dsa-tiny-random-bf16 -- published by malaiwah. Tokenizer fixture-tokenizer-08d3f8f92d3b9086, vocabulary 260.

Panel: Synthetic CPU fixture panel 525cb6c62509

Panel disclosure -- reduced_run_count: Receipt disclosure reduced_run_count: this dataset is ONE cold capture (run_count 1, the DET-D4 floor is 5). Cross-run determinism is not asserted by the dataset; it is established by a second cold capture plus fidelity-dataset compare --self-compare, whose exactly-0.0 result is the SC-1 reproduction confirmation.

Panel disclosure -- native_head_replay: Receipt disclosure native_head_replay: HEAD-1d: each side replayed through its own sealed head (reference 2364995d4009, candidate 2364995d4009); head error is inside the measurement, as under HEAD-2, and nothing is substituted. The heads are content-identical.

This panel carries 2 separate comparability groups. They are different measurements of different things and are never merged.

Group cmp--9b87b6d52c97e8ca -- 1 row

Panel panel--fixture.9d187d1dab31575b41ab4b00 -- Synthetic CPU fixture panel 525cb6c62509 4 contexts x 63 scored positions = 252 scored positions, score_from 0 sealed: yes (token digest 525cb6c625096ddb...) -- contamination scan: NOT RUN Reference (teacher) reference--fixture.glm-moe-dsa.29e66b3afcf292da -- native_bf16, artifact artifact--malaiwah.glm-moe-dsa-tiny-random-bf16.a1a973330e4e @a1a973330e4e912678adbd8ea41e8b86b1c9dccb Metric mean_tokenwise_kld, direction reference_to_candidate, accumulation float64 Estimation surface stack_relation same_stack, head_policy native_head Comparability key cmp--9b87b6d52c97e8ca Like-for-like predicate comparable: unknown -- no recorded difference, but hardware, replay_backend, replay_env, stack are unrecorded for at least one member, so homogeneity cannot be certified. Machine-readable form with per-dimension values: this key's comparability block in index.json.

What this table is. Every row here shares the comparability key above: the same tokens, the same teacher capture, the same metric and direction, the same estimator precision, the same stack relation and the same head policy. That makes them CANDIDATES for ranking -- the key is a necessary partition, not a certificate. Whether they are also like-for-like on the dimensions the key omits (lane, pipeline, scope coverage, hardware) is what the predicate line above answers.

Rank is not a verdict. The table is sorted by fidelity alone, and fidelity buys bits: a larger, higher-bitrate quant will usually sit above a smaller one, which is not news. Read the Size and Codec columns before reading the order, and compare like against like.

What it is NOT comparable to. The nearest neighbouring groups differ in:

• cmp--298751ef53a01889 (1 row): reference_id reference--fixture.glm-moe-dsa.29e66b3afcf292da -> reference--native.26e9fe0e0ea53ae8a7e1e630

Those numbers are in this file, under their own headings. Quoting one under the other heading is the mistake this layout exists to prevent: the key is a function of the panel, the teacher, the metric, the direction and the estimator, and the validator recomputes it from those fields rather than trusting the stamped value. What that catches is a row filed under a key its own fields do not produce. It does not catch a number attributed to the wrong panel in the first place -- no offline checker can. That is what the receipt digests on every row are for.

Also, and always: every table for a different model. A KL number is a divergence over one model's own vocabulary against that model's own teacher. It is not a quality score that can be carried between models.

Single-row group. This number has nothing in the registry to be ranked against. It is a stated fact, not a placing.

1 of this group's 1 rows came off a different measurement lane (other) and are tabled on their own below, not mixed into the ordering here. The key does not carry the lane; this file does.

Lane other -- 1 of this group's 1 rows

A different lane. Same key, and that is exactly the problem this table solves. The comparability key is a function of the panel, the teacher, the metric, the direction, the estimator precision, the stack relation and the head policy -- and these rows match the table above on all seven. What they do not share is the machine and the code path that produced the candidate logits, and lanes are not interchangeable. Sorting them into one list would read as a ranking; where the same artifact appears in both, it is one set of weights measured twice, not two quants.

No bridge to the sealed lane is recorded for this lane.

Artifact	Codec	Size	mean_tokenwise_kld (nats)	CI95	Top-1	Runs	Attribution	Receipt
glm-moe-dsa-tiny-random-bf16 (random test fixture) native BF16 (measurement floor)	bf16	0.0 GB	0	--	100.00 %	2 runs, bitwise identical	measured by us	receipt

Bias on glm-moe-dsa-tiny-random-bf16 (random test fixture) native BF16 -- other, direction unknown. This is the fixture's local CPU reproduction floor. Its offset against a production GPU lane was not measured; no cross-lane equivalence is claimed.

Disclosures for the rows above (3)

• fixture.glm-moe-dsa.floor.3fec5c1bbb4b9847 reduced_run_count: Receipt disclosure reduced_run_count: this dataset is ONE cold capture (run_count 1, the DET-D4 floor is 5). Cross-run determinism is not asserted by the dataset; it is established by a second cold capture plus fidelity-dataset compare --self-compare, whose exactly-0.0 result is the SC-1 reproduction confirmation.
• fixture.glm-moe-dsa.floor.3fec5c1bbb4b9847 non_sealed_lane: Local CPU fixture lane; this exact same-lane reproduction is not a measured offset against a production sealed GPU lane.
• fixture.glm-moe-dsa.floor.3fec5c1bbb4b9847 reduced_run_count: Two qualified cold runs, not five.

Group cmp--298751ef53a01889 -- 1 row

Panel panel--fixture.9d187d1dab31575b41ab4b00 -- Synthetic CPU fixture panel 525cb6c62509 4 contexts x 63 scored positions = 252 scored positions, score_from 0 sealed: yes (token digest 525cb6c625096ddb...) -- contamination scan: NOT RUN Reference (teacher) reference--native.26e9fe0e0ea53ae8a7e1e630 -- native_bf16, artifact artifact--malaiwah.glm-moe-dsa-tiny-random-bf16.a1a973330e4e @a1a973330e4e912678adbd8ea41e8b86b1c9dccb Metric mean_tokenwise_kld, direction reference_to_candidate, accumulation float64 Estimation surface stack_relation same_stack, head_policy native_head Comparability key cmp--298751ef53a01889 Like-for-like predicate comparable: unknown -- no recorded difference, but hardware, replay_backend, replay_env, stack are unrecorded for at least one member, so homogeneity cannot be certified. Machine-readable form with per-dimension values: this key's comparability block in index.json.

What this table is. Every row here shares the comparability key above: the same tokens, the same teacher capture, the same metric and direction, the same estimator precision, the same stack relation and the same head policy. That makes them CANDIDATES for ranking -- the key is a necessary partition, not a certificate. Whether they are also like-for-like on the dimensions the key omits (lane, pipeline, scope coverage, hardware) is what the predicate line above answers.

Rank is not a verdict. The table is sorted by fidelity alone, and fidelity buys bits: a larger, higher-bitrate quant will usually sit above a smaller one, which is not news. Read the Size and Codec columns before reading the order, and compare like against like.

What it is NOT comparable to. The nearest neighbouring groups differ in:

• cmp--9b87b6d52c97e8ca (1 row): reference_id reference--native.26e9fe0e0ea53ae8a7e1e630 -> reference--fixture.glm-moe-dsa.29e66b3afcf292da

Those numbers are in this file, under their own headings. Quoting one under the other heading is the mistake this layout exists to prevent: the key is a function of the panel, the teacher, the metric, the direction and the estimator, and the validator recomputes it from those fields rather than trusting the stamped value. What that catches is a row filed under a key its own fields do not produce. It does not catch a number attributed to the wrong panel in the first place -- no offline checker can. That is what the receipt digests on every row are for.

Also, and always: every table for a different model. A KL number is a divergence over one model's own vocabulary against that model's own teacher. It is not a quality score that can be carried between models.

Single-row group. This number has nothing in the registry to be ranked against. It is a stated fact, not a placing.

1 of this group's 1 rows came off a different measurement lane (other) and are tabled on their own below, not mixed into the ordering here. The key does not carry the lane; this file does.

Lane other -- 1 of this group's 1 rows

A different lane. Same key, and that is exactly the problem this table solves. The comparability key is a function of the panel, the teacher, the metric, the direction, the estimator precision, the stack relation and the head policy -- and these rows match the table above on all seven. What they do not share is the machine and the code path that produced the candidate logits, and lanes are not interchangeable. Sorting them into one list would read as a ranking; where the same artifact appears in both, it is one set of weights measured twice, not two quants.

No bridge to the sealed lane is recorded for this lane.

Artifact	Codec	Size	mean_tokenwise_kld (nats)	CI95	Top-1	Runs	Attribution	Receipt
glm-moe-dsa-tiny-random-bf16 (random test fixture) native BF16 (measurement floor)	bf16	0.0 GB	0	--	100.00 %	2 runs, bitwise identical	measured by us	receipt

Bias on glm-moe-dsa-tiny-random-bf16 (random test fixture) native BF16 -- other, direction unknown. Exact same-lane native reproduction; its offset against other lanes was not measured.

Disclosures for the rows above (4)

• native.floor.827fb81ed36b5dd9ff6e12ae reduced_run_count: reduced_run_count: this dataset is ONE cold capture (run_count 1, the DET-D4 floor is 5). Cross-run determinism is not asserted by the dataset; it is established by a second cold capture plus fidelity-dataset compare --self-compare, whose exactly-0.0 result is the SC-1 reproduction confirmation.
• native.floor.827fb81ed36b5dd9ff6e12ae reduced_run_count: Two qualified cold runs, not five.
• native.floor.827fb81ed36b5dd9ff6e12ae non_sealed_lane: This native reproduction is confined to its recorded HF Jobs lane; no production-lane equivalence is established.
• native.floor.827fb81ed36b5dd9ff6e12ae record_note: This service read authenticated HF provider metadata and validated recovered evidence; it did not independently reproduce the model run.

glm5-next-tiny-random-bf16 (random test fixture)

model--malaiwah.glm5-next-tiny-random-bf16 -- published by malaiwah. Tokenizer fixture-tokenizer-1cd570eb10aee646, vocabulary 266.

Panel: Synthetic CPU fixture panel 1ffa88108bda

Panel disclosure -- reduced_run_count: Receipt disclosure reduced_run_count: this dataset is ONE cold capture (run_count 1, the DET-D4 floor is 5). Cross-run determinism is not asserted by the dataset; it is established by a second cold capture plus fidelity-dataset compare --self-compare, whose exactly-0.0 result is the SC-1 reproduction confirmation.

Panel disclosure -- native_head_replay: Receipt disclosure native_head_replay: HEAD-1d: each side replayed through its own sealed head (reference e7cea31edbe2, candidate e7cea31edbe2); head error is inside the measurement, as under HEAD-2, and nothing is substituted. The heads are content-identical.

Group cmp--6757032314e38a59 -- 1 row

Panel panel--fixture.3e2777903c672346baceac98 -- Synthetic CPU fixture panel 1ffa88108bda 4 contexts x 63 scored positions = 252 scored positions, score_from 0 sealed: yes (token digest 1ffa88108bda139b...) -- contamination scan: NOT RUN Reference (teacher) reference--fixture.glm5-next.d0b33dd18e79aba0 -- native_bf16, artifact artifact--malaiwah.glm5-next-tiny-random-bf16.4c31348e3beb @4c31348e3beb1a8a6bd73d055464362953d768ea Metric mean_tokenwise_kld, direction reference_to_candidate, accumulation float64 Estimation surface stack_relation same_stack, head_policy native_head Comparability key cmp--6757032314e38a59 Like-for-like predicate comparable: unknown -- no recorded difference, but hardware, replay_backend, replay_env, stack are unrecorded for at least one member, so homogeneity cannot be certified. Machine-readable form with per-dimension values: this key's comparability block in index.json.

What this table is. Every row here shares the comparability key above: the same tokens, the same teacher capture, the same metric and direction, the same estimator precision, the same stack relation and the same head policy. That makes them CANDIDATES for ranking -- the key is a necessary partition, not a certificate. Whether they are also like-for-like on the dimensions the key omits (lane, pipeline, scope coverage, hardware) is what the predicate line above answers.

Rank is not a verdict. The table is sorted by fidelity alone, and fidelity buys bits: a larger, higher-bitrate quant will usually sit above a smaller one, which is not news. Read the Size and Codec columns before reading the order, and compare like against like.

What it is NOT comparable to. Every other table in this file: no other group shares this key. That includes every table for a different model -- a KL number is a divergence over one model's own vocabulary against that model's own teacher, never a score that can be carried between models.

Single-row group. This number has nothing in the registry to be ranked against. It is a stated fact, not a placing.

1 of this group's 1 rows came off a different measurement lane (other) and are tabled on their own below, not mixed into the ordering here. The key does not carry the lane; this file does.

Lane other -- 1 of this group's 1 rows

A different lane. Same key, and that is exactly the problem this table solves. The comparability key is a function of the panel, the teacher, the metric, the direction, the estimator precision, the stack relation and the head policy -- and these rows match the table above on all seven. What they do not share is the machine and the code path that produced the candidate logits, and lanes are not interchangeable. Sorting them into one list would read as a ranking; where the same artifact appears in both, it is one set of weights measured twice, not two quants.

No bridge to the sealed lane is recorded for this lane.

Artifact	Codec	Size	mean_tokenwise_kld (nats)	CI95	Top-1	Runs	Attribution	Receipt
glm5-next-tiny-random-bf16 (random test fixture) native BF16 (measurement floor)	bf16	0.0 GB	0	--	100.00 %	2 runs, bitwise identical	measured by us	receipt

Bias on glm5-next-tiny-random-bf16 (random test fixture) native BF16 -- other, direction unknown. This is the fixture's local CPU reproduction floor. Its offset against a production GPU lane was not measured; no cross-lane equivalence is claimed.

Disclosures for the rows above (3)

• fixture.glm5-next.floor.9b37d404b8b10233 reduced_run_count: Receipt disclosure reduced_run_count: this dataset is ONE cold capture (run_count 1, the DET-D4 floor is 5). Cross-run determinism is not asserted by the dataset; it is established by a second cold capture plus fidelity-dataset compare --self-compare, whose exactly-0.0 result is the SC-1 reproduction confirmation.
• fixture.glm5-next.floor.9b37d404b8b10233 non_sealed_lane: Local CPU fixture lane; this exact same-lane reproduction is not a measured offset against a production sealed GPU lane.
• fixture.glm5-next.floor.9b37d404b8b10233 reduced_run_count: Two qualified cold runs, not five.

k2-horizon-tiny-random-bf16 (random test fixture)

model--malaiwah.k2-horizon-tiny-random-bf16 -- published by malaiwah. Tokenizer fixture-tokenizer-a411cee2863e41c1, vocabulary 272.

Panel: Synthetic CPU fixture panel 525cb6c62509

Panel disclosure -- reduced_run_count: Receipt disclosure reduced_run_count: this dataset is ONE cold capture (run_count 1, the DET-D4 floor is 5). Cross-run determinism is not asserted by the dataset; it is established by a second cold capture plus fidelity-dataset compare --self-compare, whose exactly-0.0 result is the SC-1 reproduction confirmation.

Panel disclosure -- native_head_replay: Receipt disclosure native_head_replay: HEAD-1d: each side replayed through its own sealed head (reference 5c5d4b4405a5, candidate 5c5d4b4405a5); head error is inside the measurement, as under HEAD-2, and nothing is substituted. The heads are content-identical.

Group cmp--1fc318e7eb2a6ed5 -- 1 row

Panel panel--fixture.53326a9cd31fede1ea932442 -- Synthetic CPU fixture panel 525cb6c62509 4 contexts x 63 scored positions = 252 scored positions, score_from 0 sealed: yes (token digest 525cb6c625096ddb...) -- contamination scan: NOT RUN Reference (teacher) reference--fixture.k2-horizon.baeeb09002899e7b -- native_bf16, artifact artifact--malaiwah.k2-horizon-tiny-random-bf16.50627d8f2166 @50627d8f2166ee04a172f4a5144dc0b3385b9ee9 Metric mean_tokenwise_kld, direction reference_to_candidate, accumulation float64 Estimation surface stack_relation same_stack, head_policy native_head Comparability key cmp--1fc318e7eb2a6ed5 Like-for-like predicate comparable: unknown -- no recorded difference, but hardware, replay_backend, replay_env, stack are unrecorded for at least one member, so homogeneity cannot be certified. Machine-readable form with per-dimension values: this key's comparability block in index.json.

What this table is. Every row here shares the comparability key above: the same tokens, the same teacher capture, the same metric and direction, the same estimator precision, the same stack relation and the same head policy. That makes them CANDIDATES for ranking -- the key is a necessary partition, not a certificate. Whether they are also like-for-like on the dimensions the key omits (lane, pipeline, scope coverage, hardware) is what the predicate line above answers.

Rank is not a verdict. The table is sorted by fidelity alone, and fidelity buys bits: a larger, higher-bitrate quant will usually sit above a smaller one, which is not news. Read the Size and Codec columns before reading the order, and compare like against like.

What it is NOT comparable to. Every other table in this file: no other group shares this key. That includes every table for a different model -- a KL number is a divergence over one model's own vocabulary against that model's own teacher, never a score that can be carried between models.

Single-row group. This number has nothing in the registry to be ranked against. It is a stated fact, not a placing.

1 of this group's 1 rows came off a different measurement lane (other) and are tabled on their own below, not mixed into the ordering here. The key does not carry the lane; this file does.

Lane other -- 1 of this group's 1 rows

A different lane. Same key, and that is exactly the problem this table solves. The comparability key is a function of the panel, the teacher, the metric, the direction, the estimator precision, the stack relation and the head policy -- and these rows match the table above on all seven. What they do not share is the machine and the code path that produced the candidate logits, and lanes are not interchangeable. Sorting them into one list would read as a ranking; where the same artifact appears in both, it is one set of weights measured twice, not two quants.

No bridge to the sealed lane is recorded for this lane.

Artifact	Codec	Size	mean_tokenwise_kld (nats)	CI95	Top-1	Runs	Attribution	Receipt
k2-horizon-tiny-random-bf16 (random test fixture) native BF16 (measurement floor)	bf16	0.0 GB	0	--	100.00 %	2 runs, bitwise identical	measured by us	receipt

Bias on k2-horizon-tiny-random-bf16 (random test fixture) native BF16 -- other, direction unknown. This is the fixture's local CPU reproduction floor. Its offset against a production GPU lane was not measured; no cross-lane equivalence is claimed.

Disclosures for the rows above (3)

• fixture.k2-horizon.floor.bb50aed4c5b1f157 reduced_run_count: Receipt disclosure reduced_run_count: this dataset is ONE cold capture (run_count 1, the DET-D4 floor is 5). Cross-run determinism is not asserted by the dataset; it is established by a second cold capture plus fidelity-dataset compare --self-compare, whose exactly-0.0 result is the SC-1 reproduction confirmation.
• fixture.k2-horizon.floor.bb50aed4c5b1f157 non_sealed_lane: Local CPU fixture lane; this exact same-lane reproduction is not a measured offset against a production sealed GPU lane.
• fixture.k2-horizon.floor.bb50aed4c5b1f157 reduced_run_count: Two qualified cold runs, not five.

kimi-k25-tiny-random-bf16 (random test fixture)

model--malaiwah.kimi-k25-tiny-random-bf16 -- published by malaiwah. Tokenizer fixture-tokenizer-9cc132a24ced4fd7, vocabulary 272.

Panel: Synthetic CPU fixture panel 525cb6c62509

Panel disclosure -- reduced_run_count: Receipt disclosure reduced_run_count: this dataset is ONE cold capture (run_count 1, the DET-D4 floor is 5). Cross-run determinism is not asserted by the dataset; it is established by a second cold capture plus fidelity-dataset compare --self-compare, whose exactly-0.0 result is the SC-1 reproduction confirmation.

Panel disclosure -- native_head_replay: Receipt disclosure native_head_replay: HEAD-1d: each side replayed through its own sealed head (reference e68a813f441b, candidate e68a813f441b); head error is inside the measurement, as under HEAD-2, and nothing is substituted. The heads are content-identical.

Group cmp--2cd2acab7abc566a -- 1 row

Panel panel--fixture.a53145fb225cc15b73887d0f -- Synthetic CPU fixture panel 525cb6c62509 4 contexts x 63 scored positions = 252 scored positions, score_from 0 sealed: yes (token digest 525cb6c625096ddb...) -- contamination scan: NOT RUN Reference (teacher) reference--fixture.kimi-k25.29336d88a8aa2634 -- native_bf16, artifact artifact--malaiwah.kimi-k25-tiny-random-bf16.721c97dd956f @721c97dd956f23b6b5c1040d633539a162a19094 Metric mean_tokenwise_kld, direction reference_to_candidate, accumulation float64 Estimation surface stack_relation same_stack, head_policy native_head Comparability key cmp--2cd2acab7abc566a Like-for-like predicate comparable: unknown -- no recorded difference, but hardware, replay_backend, replay_env, stack are unrecorded for at least one member, so homogeneity cannot be certified. Machine-readable form with per-dimension values: this key's comparability block in index.json.

What this table is. Every row here shares the comparability key above: the same tokens, the same teacher capture, the same metric and direction, the same estimator precision, the same stack relation and the same head policy. That makes them CANDIDATES for ranking -- the key is a necessary partition, not a certificate. Whether they are also like-for-like on the dimensions the key omits (lane, pipeline, scope coverage, hardware) is what the predicate line above answers.

Rank is not a verdict. The table is sorted by fidelity alone, and fidelity buys bits: a larger, higher-bitrate quant will usually sit above a smaller one, which is not news. Read the Size and Codec columns before reading the order, and compare like against like.

What it is NOT comparable to. Every other table in this file: no other group shares this key. That includes every table for a different model -- a KL number is a divergence over one model's own vocabulary against that model's own teacher, never a score that can be carried between models.

Single-row group. This number has nothing in the registry to be ranked against. It is a stated fact, not a placing.

1 of this group's 1 rows came off a different measurement lane (other) and are tabled on their own below, not mixed into the ordering here. The key does not carry the lane; this file does.

Lane other -- 1 of this group's 1 rows

A different lane. Same key, and that is exactly the problem this table solves. The comparability key is a function of the panel, the teacher, the metric, the direction, the estimator precision, the stack relation and the head policy -- and these rows match the table above on all seven. What they do not share is the machine and the code path that produced the candidate logits, and lanes are not interchangeable. Sorting them into one list would read as a ranking; where the same artifact appears in both, it is one set of weights measured twice, not two quants.

No bridge to the sealed lane is recorded for this lane.

Artifact	Codec	Size	mean_tokenwise_kld (nats)	CI95	Top-1	Runs	Attribution	Receipt
kimi-k25-tiny-random-bf16 (random test fixture) native BF16 (measurement floor)	bf16	0.0 GB	0	--	100.00 %	2 runs, bitwise identical	measured by us	receipt

Bias on kimi-k25-tiny-random-bf16 (random test fixture) native BF16 -- other, direction unknown. This is the fixture's local CPU reproduction floor. Its offset against a production GPU lane was not measured; no cross-lane equivalence is claimed.

Disclosures for the rows above (3)

• fixture.kimi-k25.floor.96968fae6bcf3237 reduced_run_count: Receipt disclosure reduced_run_count: this dataset is ONE cold capture (run_count 1, the DET-D4 floor is 5). Cross-run determinism is not asserted by the dataset; it is established by a second cold capture plus fidelity-dataset compare --self-compare, whose exactly-0.0 result is the SC-1 reproduction confirmation.
• fixture.kimi-k25.floor.96968fae6bcf3237 non_sealed_lane: Local CPU fixture lane; this exact same-lane reproduction is not a measured offset against a production sealed GPU lane.
• fixture.kimi-k25.floor.96968fae6bcf3237 reduced_run_count: Two qualified cold runs, not five.

kimi-k3-tiny-random-bf16 (random test fixture)

model--malaiwah.kimi-k3-tiny-random-bf16 -- published by malaiwah. Tokenizer fixture-tokenizer-9cc132a24ced4fd7, vocabulary 272.

Panel: Synthetic CPU fixture panel 525cb6c62509

Panel disclosure -- reduced_run_count: Receipt disclosure reduced_run_count: this dataset is ONE cold capture (run_count 1, the DET-D4 floor is 5). Cross-run determinism is not asserted by the dataset; it is established by a second cold capture plus fidelity-dataset compare --self-compare, whose exactly-0.0 result is the SC-1 reproduction confirmation.

Panel disclosure -- native_head_replay: Receipt disclosure native_head_replay: HEAD-1d: each side replayed through its own sealed head (reference e68a813f441b, candidate e68a813f441b); head error is inside the measurement, as under HEAD-2, and nothing is substituted. The heads are content-identical.

Group cmp--4a66bc7afad67fe2 -- 1 row

Panel panel--fixture.a53145fb225cc15b73887d0f -- Synthetic CPU fixture panel 525cb6c62509 4 contexts x 63 scored positions = 252 scored positions, score_from 0 sealed: yes (token digest 525cb6c625096ddb...) -- contamination scan: NOT RUN Reference (teacher) reference--fixture.kimi-k3.c90350e92d0b404a -- native_bf16, artifact artifact--malaiwah.kimi-k3-tiny-random-bf16.3c1e2f22e5cc @3c1e2f22e5ccfe37addc9906bfb0fdea6ad00e16 Metric mean_tokenwise_kld, direction reference_to_candidate, accumulation float64 Estimation surface stack_relation same_stack, head_policy native_head Comparability key cmp--4a66bc7afad67fe2 Like-for-like predicate comparable: unknown -- no recorded difference, but hardware, replay_backend, replay_env, stack are unrecorded for at least one member, so homogeneity cannot be certified. Machine-readable form with per-dimension values: this key's comparability block in index.json.

What this table is. Every row here shares the comparability key above: the same tokens, the same teacher capture, the same metric and direction, the same estimator precision, the same stack relation and the same head policy. That makes them CANDIDATES for ranking -- the key is a necessary partition, not a certificate. Whether they are also like-for-like on the dimensions the key omits (lane, pipeline, scope coverage, hardware) is what the predicate line above answers.

Rank is not a verdict. The table is sorted by fidelity alone, and fidelity buys bits: a larger, higher-bitrate quant will usually sit above a smaller one, which is not news. Read the Size and Codec columns before reading the order, and compare like against like.

What it is NOT comparable to. Every other table in this file: no other group shares this key. That includes every table for a different model -- a KL number is a divergence over one model's own vocabulary against that model's own teacher, never a score that can be carried between models.

Single-row group. This number has nothing in the registry to be ranked against. It is a stated fact, not a placing.

1 of this group's 1 rows came off a different measurement lane (other) and are tabled on their own below, not mixed into the ordering here. The key does not carry the lane; this file does.

Lane other -- 1 of this group's 1 rows

A different lane. Same key, and that is exactly the problem this table solves. The comparability key is a function of the panel, the teacher, the metric, the direction, the estimator precision, the stack relation and the head policy -- and these rows match the table above on all seven. What they do not share is the machine and the code path that produced the candidate logits, and lanes are not interchangeable. Sorting them into one list would read as a ranking; where the same artifact appears in both, it is one set of weights measured twice, not two quants.

No bridge to the sealed lane is recorded for this lane.

Artifact	Codec	Size	mean_tokenwise_kld (nats)	CI95	Top-1	Runs	Attribution	Receipt
kimi-k3-tiny-random-bf16 (random test fixture) native BF16 (measurement floor)	bf16	0.0 GB	0	--	100.00 %	2 runs, bitwise identical	measured by us	receipt

Bias on kimi-k3-tiny-random-bf16 (random test fixture) native BF16 -- other, direction unknown. This is the fixture's local CPU reproduction floor. Its offset against a production GPU lane was not measured; no cross-lane equivalence is claimed.

Disclosures for the rows above (3)

• fixture.kimi-k3.floor.2e3a2daddd930f9a reduced_run_count: Receipt disclosure reduced_run_count: this dataset is ONE cold capture (run_count 1, the DET-D4 floor is 5). Cross-run determinism is not asserted by the dataset; it is established by a second cold capture plus fidelity-dataset compare --self-compare, whose exactly-0.0 result is the SC-1 reproduction confirmation.
• fixture.kimi-k3.floor.2e3a2daddd930f9a non_sealed_lane: Local CPU fixture lane; this exact same-lane reproduction is not a measured offset against a production sealed GPU lane.
• fixture.kimi-k3.floor.2e3a2daddd930f9a reduced_run_count: Two qualified cold runs, not five.

minimax-m2-tiny-random-bf16 (random test fixture)

model--malaiwah.minimax-m2-tiny-random-bf16 -- published by malaiwah. Tokenizer fixture-tokenizer-9cc132a24ced4fd7, vocabulary 272.

Panel: Synthetic CPU fixture panel 525cb6c62509

Panel disclosure -- reduced_run_count: Receipt disclosure reduced_run_count: this dataset is ONE cold capture (run_count 1, the DET-D4 floor is 5). Cross-run determinism is not asserted by the dataset; it is established by a second cold capture plus fidelity-dataset compare --self-compare, whose exactly-0.0 result is the SC-1 reproduction confirmation.

Panel disclosure -- native_head_replay: Receipt disclosure native_head_replay: HEAD-1d: each side replayed through its own sealed head (reference e68a813f441b, candidate e68a813f441b); head error is inside the measurement, as under HEAD-2, and nothing is substituted. The heads are content-identical.

Group cmp--a99c67940f387a86 -- 1 row

Panel panel--fixture.a53145fb225cc15b73887d0f -- Synthetic CPU fixture panel 525cb6c62509 4 contexts x 63 scored positions = 252 scored positions, score_from 0 sealed: yes (token digest 525cb6c625096ddb...) -- contamination scan: NOT RUN Reference (teacher) reference--fixture.minimax-m2.9a0b71a3347e072b -- native_bf16, artifact artifact--malaiwah.minimax-m2-tiny-random-bf16.d4825603496f @d4825603496f1c636385af403e891aaf20e9afe6 Metric mean_tokenwise_kld, direction reference_to_candidate, accumulation float64 Estimation surface stack_relation same_stack, head_policy native_head Comparability key cmp--a99c67940f387a86 Like-for-like predicate comparable: unknown -- no recorded difference, but hardware, replay_backend, replay_env, stack are unrecorded for at least one member, so homogeneity cannot be certified. Machine-readable form with per-dimension values: this key's comparability block in index.json.

What this table is. Every row here shares the comparability key above: the same tokens, the same teacher capture, the same metric and direction, the same estimator precision, the same stack relation and the same head policy. That makes them CANDIDATES for ranking -- the key is a necessary partition, not a certificate. Whether they are also like-for-like on the dimensions the key omits (lane, pipeline, scope coverage, hardware) is what the predicate line above answers.

Rank is not a verdict. The table is sorted by fidelity alone, and fidelity buys bits: a larger, higher-bitrate quant will usually sit above a smaller one, which is not news. Read the Size and Codec columns before reading the order, and compare like against like.

What it is NOT comparable to. Every other table in this file: no other group shares this key. That includes every table for a different model -- a KL number is a divergence over one model's own vocabulary against that model's own teacher, never a score that can be carried between models.

Single-row group. This number has nothing in the registry to be ranked against. It is a stated fact, not a placing.

1 of this group's 1 rows came off a different measurement lane (other) and are tabled on their own below, not mixed into the ordering here. The key does not carry the lane; this file does.

Lane other -- 1 of this group's 1 rows

A different lane. Same key, and that is exactly the problem this table solves. The comparability key is a function of the panel, the teacher, the metric, the direction, the estimator precision, the stack relation and the head policy -- and these rows match the table above on all seven. What they do not share is the machine and the code path that produced the candidate logits, and lanes are not interchangeable. Sorting them into one list would read as a ranking; where the same artifact appears in both, it is one set of weights measured twice, not two quants.

No bridge to the sealed lane is recorded for this lane.

Artifact	Codec	Size	mean_tokenwise_kld (nats)	CI95	Top-1	Runs	Attribution	Receipt
minimax-m2-tiny-random-bf16 (random test fixture) native BF16 (measurement floor)	bf16	0.0 GB	0	--	100.00 %	2 runs, bitwise identical	measured by us	receipt

Bias on minimax-m2-tiny-random-bf16 (random test fixture) native BF16 -- other, direction unknown. This is the fixture's local CPU reproduction floor. Its offset against a production GPU lane was not measured; no cross-lane equivalence is claimed.

Disclosures for the rows above (3)

• fixture.minimax-m2.floor.ca1e98dca763d440 reduced_run_count: Receipt disclosure reduced_run_count: this dataset is ONE cold capture (run_count 1, the DET-D4 floor is 5). Cross-run determinism is not asserted by the dataset; it is established by a second cold capture plus fidelity-dataset compare --self-compare, whose exactly-0.0 result is the SC-1 reproduction confirmation.
• fixture.minimax-m2.floor.ca1e98dca763d440 non_sealed_lane: Local CPU fixture lane; this exact same-lane reproduction is not a measured offset against a production sealed GPU lane.
• fixture.minimax-m2.floor.ca1e98dca763d440 reduced_run_count: Two qualified cold runs, not five.

minimax-m3-tiny-random-bf16 (random test fixture)

model--malaiwah.minimax-m3-tiny-random-bf16 -- published by malaiwah. Tokenizer fixture-tokenizer-9cc132a24ced4fd7, vocabulary 272.

Panel: Synthetic CPU fixture panel 525cb6c62509

Panel disclosure -- reduced_run_count: Receipt disclosure reduced_run_count: this dataset is ONE cold capture (run_count 1, the DET-D4 floor is 5). Cross-run determinism is not asserted by the dataset; it is established by a second cold capture plus fidelity-dataset compare --self-compare, whose exactly-0.0 result is the SC-1 reproduction confirmation.

Panel disclosure -- native_head_replay: Receipt disclosure native_head_replay: HEAD-1d: each side replayed through its own sealed head (reference e68a813f441b, candidate e68a813f441b); head error is inside the measurement, as under HEAD-2, and nothing is substituted. The heads are content-identical.

Group cmp--7e10ab9bcc7628b7 -- 1 row

Panel panel--fixture.a53145fb225cc15b73887d0f -- Synthetic CPU fixture panel 525cb6c62509 4 contexts x 63 scored positions = 252 scored positions, score_from 0 sealed: yes (token digest 525cb6c625096ddb...) -- contamination scan: NOT RUN Reference (teacher) reference--fixture.minimax-m3.d193c0296d88fa50 -- native_bf16, artifact artifact--malaiwah.minimax-m3-tiny-random-bf16.ca904dff2941 @ca904dff2941004a3f5bd5af62064fcfb34da8b9 Metric mean_tokenwise_kld, direction reference_to_candidate, accumulation float64 Estimation surface stack_relation same_stack, head_policy native_head Comparability key cmp--7e10ab9bcc7628b7 Like-for-like predicate comparable: unknown -- no recorded difference, but hardware, replay_backend, replay_env, stack are unrecorded for at least one member, so homogeneity cannot be certified. Machine-readable form with per-dimension values: this key's comparability block in index.json.

What this table is. Every row here shares the comparability key above: the same tokens, the same teacher capture, the same metric and direction, the same estimator precision, the same stack relation and the same head policy. That makes them CANDIDATES for ranking -- the key is a necessary partition, not a certificate. Whether they are also like-for-like on the dimensions the key omits (lane, pipeline, scope coverage, hardware) is what the predicate line above answers.

Rank is not a verdict. The table is sorted by fidelity alone, and fidelity buys bits: a larger, higher-bitrate quant will usually sit above a smaller one, which is not news. Read the Size and Codec columns before reading the order, and compare like against like.

What it is NOT comparable to. Every other table in this file: no other group shares this key. That includes every table for a different model -- a KL number is a divergence over one model's own vocabulary against that model's own teacher, never a score that can be carried between models.

Single-row group. This number has nothing in the registry to be ranked against. It is a stated fact, not a placing.

1 of this group's 1 rows came off a different measurement lane (other) and are tabled on their own below, not mixed into the ordering here. The key does not carry the lane; this file does.

Lane other -- 1 of this group's 1 rows

A different lane. Same key, and that is exactly the problem this table solves. The comparability key is a function of the panel, the teacher, the metric, the direction, the estimator precision, the stack relation and the head policy -- and these rows match the table above on all seven. What they do not share is the machine and the code path that produced the candidate logits, and lanes are not interchangeable. Sorting them into one list would read as a ranking; where the same artifact appears in both, it is one set of weights measured twice, not two quants.

No bridge to the sealed lane is recorded for this lane.

Artifact	Codec	Size	mean_tokenwise_kld (nats)	CI95	Top-1	Runs	Attribution	Receipt
minimax-m3-tiny-random-bf16 (random test fixture) native BF16 (measurement floor)	bf16	0.0 GB	0	--	100.00 %	2 runs, bitwise identical	measured by us	receipt

Bias on minimax-m3-tiny-random-bf16 (random test fixture) native BF16 -- other, direction unknown. This is the fixture's local CPU reproduction floor. Its offset against a production GPU lane was not measured; no cross-lane equivalence is claimed.

Disclosures for the rows above (3)

• fixture.minimax-m3.floor.45b1ac062143d1d7 reduced_run_count: Receipt disclosure reduced_run_count: this dataset is ONE cold capture (run_count 1, the DET-D4 floor is 5). Cross-run determinism is not asserted by the dataset; it is established by a second cold capture plus fidelity-dataset compare --self-compare, whose exactly-0.0 result is the SC-1 reproduction confirmation.
• fixture.minimax-m3.floor.45b1ac062143d1d7 non_sealed_lane: Local CPU fixture lane; this exact same-lane reproduction is not a measured offset against a production sealed GPU lane.
• fixture.minimax-m3.floor.45b1ac062143d1d7 reduced_run_count: Two qualified cold runs, not five.

qwen3-5-gguf-tiny-random-bf16 (random test fixture)

model--malaiwah.qwen3-5-gguf-tiny-random-bf16 -- published by malaiwah. Tokenizer fixture-tokenizer-9cc132a24ced4fd7, vocabulary 272.

Panel: Synthetic CPU fixture panel 525cb6c62509

Panel disclosure -- reduced_run_count: Receipt disclosure reduced_run_count: this dataset is ONE cold capture (run_count 1, the DET-D4 floor is 5). Cross-run determinism is not asserted by the dataset; it is established by a second cold capture plus fidelity-dataset compare --self-compare, whose exactly-0.0 result is the SC-1 reproduction confirmation.

Panel disclosure -- native_head_replay: Receipt disclosure native_head_replay: HEAD-1d: each side replayed through its own sealed head (reference e68a813f441b, candidate e68a813f441b); head error is inside the measurement, as under HEAD-2, and nothing is substituted. The heads are content-identical.

Group cmp--4941dc6687219470 -- 1 row

Panel panel--fixture.a53145fb225cc15b73887d0f -- Synthetic CPU fixture panel 525cb6c62509 4 contexts x 63 scored positions = 252 scored positions, score_from 0 sealed: yes (token digest 525cb6c625096ddb...) -- contamination scan: NOT RUN Reference (teacher) reference--fixture.qwen35-gguf.5976d6ef519b46f8 -- native_bf16, artifact artifact--malaiwah.qwen3-5-gguf-tiny-random-bf16.490767e58441 @490767e58441a55c5a9f6375ea0e31e2cdc9da76 Metric mean_tokenwise_kld, direction reference_to_candidate, accumulation float64 Estimation surface stack_relation same_stack, head_policy native_head Comparability key cmp--4941dc6687219470 Like-for-like predicate comparable: unknown -- no recorded difference, but hardware, replay_backend, replay_env, stack are unrecorded for at least one member, so homogeneity cannot be certified. Machine-readable form with per-dimension values: this key's comparability block in index.json.

What this table is. Every row here shares the comparability key above: the same tokens, the same teacher capture, the same metric and direction, the same estimator precision, the same stack relation and the same head policy. That makes them CANDIDATES for ranking -- the key is a necessary partition, not a certificate. Whether they are also like-for-like on the dimensions the key omits (lane, pipeline, scope coverage, hardware) is what the predicate line above answers.

Rank is not a verdict. The table is sorted by fidelity alone, and fidelity buys bits: a larger, higher-bitrate quant will usually sit above a smaller one, which is not news. Read the Size and Codec columns before reading the order, and compare like against like.

What it is NOT comparable to. Every other table in this file: no other group shares this key. That includes every table for a different model -- a KL number is a divergence over one model's own vocabulary against that model's own teacher, never a score that can be carried between models.

Single-row group. This number has nothing in the registry to be ranked against. It is a stated fact, not a placing.

1 of this group's 1 rows came off a different measurement lane (other) and are tabled on their own below, not mixed into the ordering here. The key does not carry the lane; this file does.

Lane other -- 1 of this group's 1 rows

A different lane. Same key, and that is exactly the problem this table solves. The comparability key is a function of the panel, the teacher, the metric, the direction, the estimator precision, the stack relation and the head policy -- and these rows match the table above on all seven. What they do not share is the machine and the code path that produced the candidate logits, and lanes are not interchangeable. Sorting them into one list would read as a ranking; where the same artifact appears in both, it is one set of weights measured twice, not two quants.

No bridge to the sealed lane is recorded for this lane.

Artifact	Codec	Size	mean_tokenwise_kld (nats)	CI95	Top-1	Runs	Attribution	Receipt
qwen3-5-gguf-tiny-random-bf16 (random test fixture) native BF16 (measurement floor)	bf16	0.0 GB	0	--	100.00 %	2 runs, bitwise identical	measured by us	receipt

Bias on qwen3-5-gguf-tiny-random-bf16 (random test fixture) native BF16 -- other, direction unknown. This is the fixture's local CPU reproduction floor. Its offset against a production GPU lane was not measured; no cross-lane equivalence is claimed.

Disclosures for the rows above (3)

• fixture.qwen35-gguf.floor.9b7444090207bc70 reduced_run_count: Receipt disclosure reduced_run_count: this dataset is ONE cold capture (run_count 1, the DET-D4 floor is 5). Cross-run determinism is not asserted by the dataset; it is established by a second cold capture plus fidelity-dataset compare --self-compare, whose exactly-0.0 result is the SC-1 reproduction confirmation.
• fixture.qwen35-gguf.floor.9b7444090207bc70 non_sealed_lane: Local CPU fixture lane; this exact same-lane reproduction is not a measured offset against a production sealed GPU lane.
• fixture.qwen35-gguf.floor.9b7444090207bc70 reduced_run_count: Two qualified cold runs, not five.

qwen3-5-tiny-random-bf16 (random test fixture)

model--malaiwah.qwen3-5-tiny-random-bf16 -- published by malaiwah. Tokenizer fixture-tokenizer-9cc132a24ced4fd7, vocabulary 272.

Panel: Synthetic CPU fixture panel 525cb6c62509

Panel disclosure -- reduced_run_count: Receipt disclosure reduced_run_count: this dataset is ONE cold capture (run_count 1, the DET-D4 floor is 5). Cross-run determinism is not asserted by the dataset; it is established by a second cold capture plus fidelity-dataset compare --self-compare, whose exactly-0.0 result is the SC-1 reproduction confirmation.

Panel disclosure -- native_head_replay: Receipt disclosure native_head_replay: HEAD-1d: each side replayed through its own sealed head (reference e68a813f441b, candidate e68a813f441b); head error is inside the measurement, as under HEAD-2, and nothing is substituted. The heads are content-identical.

This panel carries 2 separate comparability groups. They are different measurements of different things and are never merged.

Group cmp--238563353ac89178 -- 1 row

Panel panel--fixture.a53145fb225cc15b73887d0f -- Synthetic CPU fixture panel 525cb6c62509 4 contexts x 63 scored positions = 252 scored positions, score_from 0 sealed: yes (token digest 525cb6c625096ddb...) -- contamination scan: NOT RUN Reference (teacher) reference--fixture.qwen3-5.1847b664a523e921 -- native_bf16, artifact artifact--malaiwah.qwen3-5-tiny-random-bf16.a430e41d5814 @a430e41d5814ba5e7bfa5ee29d86935aac56d95d Metric mean_tokenwise_kld, direction reference_to_candidate, accumulation float64 Estimation surface stack_relation same_stack, head_policy native_head Comparability key cmp--238563353ac89178 Like-for-like predicate comparable: unknown -- no recorded difference, but hardware, replay_backend, replay_env, stack are unrecorded for at least one member, so homogeneity cannot be certified. Machine-readable form with per-dimension values: this key's comparability block in index.json.

What this table is. Every row here shares the comparability key above: the same tokens, the same teacher capture, the same metric and direction, the same estimator precision, the same stack relation and the same head policy. That makes them CANDIDATES for ranking -- the key is a necessary partition, not a certificate. Whether they are also like-for-like on the dimensions the key omits (lane, pipeline, scope coverage, hardware) is what the predicate line above answers.

Rank is not a verdict. The table is sorted by fidelity alone, and fidelity buys bits: a larger, higher-bitrate quant will usually sit above a smaller one, which is not news. Read the Size and Codec columns before reading the order, and compare like against like.

What it is NOT comparable to. The nearest neighbouring groups differ in:

• cmp--823cfa3f5010c0fc (1 row): stack_relation same_stack -> cross_stack

Those numbers are in this file, under their own headings. Quoting one under the other heading is the mistake this layout exists to prevent: the key is a function of the panel, the teacher, the metric, the direction and the estimator, and the validator recomputes it from those fields rather than trusting the stamped value. What that catches is a row filed under a key its own fields do not produce. It does not catch a number attributed to the wrong panel in the first place -- no offline checker can. That is what the receipt digests on every row are for.

Also, and always: every table for a different model. A KL number is a divergence over one model's own vocabulary against that model's own teacher. It is not a quality score that can be carried between models.

Single-row group. This number has nothing in the registry to be ranked against. It is a stated fact, not a placing.

1 of this group's 1 rows came off a different measurement lane (other) and are tabled on their own below, not mixed into the ordering here. The key does not carry the lane; this file does.

Lane other -- 1 of this group's 1 rows

A different lane. Same key, and that is exactly the problem this table solves. The comparability key is a function of the panel, the teacher, the metric, the direction, the estimator precision, the stack relation and the head policy -- and these rows match the table above on all seven. What they do not share is the machine and the code path that produced the candidate logits, and lanes are not interchangeable. Sorting them into one list would read as a ranking; where the same artifact appears in both, it is one set of weights measured twice, not two quants.

No bridge to the sealed lane is recorded for this lane.

Artifact	Codec	Size	mean_tokenwise_kld (nats)	CI95	Top-1	Runs	Attribution	Receipt
qwen3-5-tiny-random-bf16 (random test fixture) native BF16 (measurement floor)	bf16	0.0 GB	0	--	100.00 %	2 runs, bitwise identical	measured by us	receipt

Bias on qwen3-5-tiny-random-bf16 (random test fixture) native BF16 -- other, direction unknown. This is the fixture's local CPU reproduction floor. Its offset against a production GPU lane was not measured; no cross-lane equivalence is claimed.

Disclosures for the rows above (3)

• fixture.qwen3-5.floor.e22167048fb26e61 reduced_run_count: Receipt disclosure reduced_run_count: this dataset is ONE cold capture (run_count 1, the DET-D4 floor is 5). Cross-run determinism is not asserted by the dataset; it is established by a second cold capture plus fidelity-dataset compare --self-compare, whose exactly-0.0 result is the SC-1 reproduction confirmation.
• fixture.qwen3-5.floor.e22167048fb26e61 non_sealed_lane: Local CPU fixture lane; this exact same-lane reproduction is not a measured offset against a production sealed GPU lane.
• fixture.qwen3-5.floor.e22167048fb26e61 reduced_run_count: Two qualified cold runs, not five.

Group cmp--823cfa3f5010c0fc -- 1 row

Panel panel--fixture.a53145fb225cc15b73887d0f -- Synthetic CPU fixture panel 525cb6c62509 4 contexts x 63 scored positions = 252 scored positions, score_from 0 sealed: yes (token digest 525cb6c625096ddb...) -- contamination scan: NOT RUN Reference (teacher) reference--fixture.qwen3-5.1847b664a523e921 -- native_bf16, artifact artifact--malaiwah.qwen3-5-tiny-random-bf16.a430e41d5814 @a430e41d5814ba5e7bfa5ee29d86935aac56d95d Metric mean_tokenwise_kld, direction reference_to_candidate, accumulation float64 Estimation surface stack_relation cross_stack, head_policy native_head Comparability key cmp--823cfa3f5010c0fc Like-for-like predicate comparable: unknown -- no recorded difference, but hardware, stack are unrecorded for at least one member, so homogeneity cannot be certified. Machine-readable form with per-dimension values: this key's comparability block in index.json.

What this table is. Every row here shares the comparability key above: the same tokens, the same teacher capture, the same metric and direction, the same estimator precision, the same stack relation and the same head policy. That makes them CANDIDATES for ranking -- the key is a necessary partition, not a certificate. Whether they are also like-for-like on the dimensions the key omits (lane, pipeline, scope coverage, hardware) is what the predicate line above answers.

Rank is not a verdict. The table is sorted by fidelity alone, and fidelity buys bits: a larger, higher-bitrate quant will usually sit above a smaller one, which is not news. Read the Size and Codec columns before reading the order, and compare like against like.

What it is NOT comparable to. The nearest neighbouring groups differ in:

• cmp--238563353ac89178 (1 row): stack_relation cross_stack -> same_stack

Those numbers are in this file, under their own headings. Quoting one under the other heading is the mistake this layout exists to prevent: the key is a function of the panel, the teacher, the metric, the direction and the estimator, and the validator recomputes it from those fields rather than trusting the stamped value. What that catches is a row filed under a key its own fields do not produce. It does not catch a number attributed to the wrong panel in the first place -- no offline checker can. That is what the receipt digests on every row are for.

Also, and always: every table for a different model. A KL number is a divergence over one model's own vocabulary against that model's own teacher. It is not a quality score that can be carried between models.

Single-row group. This number has nothing in the registry to be ranked against. It is a stated fact, not a placing.

1 of this group's 1 rows came off a different measurement lane (other) and are tabled on their own below, not mixed into the ordering here. The key does not carry the lane; this file does.

Lane other -- 1 of this group's 1 rows

A different lane. Same key, and that is exactly the problem this table solves. The comparability key is a function of the panel, the teacher, the metric, the direction, the estimator precision, the stack relation and the head policy -- and these rows match the table above on all seven. What they do not share is the machine and the code path that produced the candidate logits, and lanes are not interchangeable. Sorting them into one list would read as a ranking; where the same artifact appears in both, it is one set of weights measured twice, not two quants.

No bridge to the sealed lane is recorded for this lane.

Artifact	Codec	Size	mean_tokenwise_kld (nats)	CI95	Top-1	Runs	Attribution	Receipt
qwen3-5-tiny-random-gptq-v1-g32-rtn-format (FORMAT fixture) (measurement floor)	gptq @4	0.0 GB	2.39668e-05	--	94.84 %	1 run, unevidenced	measured by us	receipt

Bias on qwen3-5-tiny-random-gptq-v1-g32-rtn-format (FORMAT fixture) -- cross_stack_capture_replay, direction unknown. the capture stacks differ or their code identity is unproven. The sign and magnitude of the cross-stack effect are unknown; a control measured on another artifact, panel or stack is not a transferable correction. usable_as_floor is false.

Disclosures for the rows above (4)

• malaiwah.qwen3-5-tiny-random-gptq-v1-g32-rtn-format.fixture.a53145fb225cc15b73887d0f cross_stack_capture: capture stack equality is not established (lane_identity equal on both sides, or unrecorded, stack_fingerprint equal on both sides, or unrecorded, source_files differ or unrecorded). The cross-stack effect has unknown sign and magnitude.
• malaiwah.qwen3-5-tiny-random-gptq-v1-g32-rtn-format.fixture.a53145fb225cc15b73887d0f weights_reconstructed: candidate was captured from a WEIGHTS-ONLY RECONSTRUCTION (affine-weight-reconstruction, output bfloat16). The stored weights were decoded before the model forward; native quantized GEMM and serving activation arithmetic are not measured. Decoder provenance is in the sealed runtime receipt. The comparison is advisory.
• malaiwah.qwen3-5-tiny-random-gptq-v1-g32-rtn-format.fixture.a53145fb225cc15b73887d0f non_sealed_lane: Produced by the 'other' lane, not the sealed-ep8 lane that the other rows in this comparability group used. Lanes are not interchangeable: this row carries an undisclosed offset against the sealed lane on the same panel until that offset is itself measured and recorded here.
• malaiwah.qwen3-5-tiny-random-gptq-v1-g32-rtn-format.fixture.a53145fb225cc15b73887d0f record_note: This service read authenticated HF provider metadata and validated recovered evidence; it did not independently reproduce the model run.

qwen4-exp-tiny-random-bf16 (random test fixture)

model--malaiwah.qwen4-exp-tiny-random-bf16 -- published by malaiwah. Tokenizer fixture-tokenizer-57630af28506f7ed, vocabulary 264.

Panel: Synthetic CPU fixture panel d0ae96d07d4a

Panel disclosure -- reduced_run_count: Receipt disclosure reduced_run_count: this dataset is ONE cold capture (run_count 1, the DET-D4 floor is 5). Cross-run determinism is not asserted by the dataset; it is established by a second cold capture plus fidelity-dataset compare --self-compare, whose exactly-0.0 result is the SC-1 reproduction confirmation.

Panel disclosure -- native_head_replay: Receipt disclosure native_head_replay: HEAD-1d: each side replayed through its own sealed head (reference f0e1210bca4f, candidate f0e1210bca4f); head error is inside the measurement, as under HEAD-2, and nothing is substituted. The heads are content-identical.

Group cmp--57fd571961121fe6 -- 1 row

Panel panel--fixture.2218037d11b09e931ba412f6 -- Synthetic CPU fixture panel d0ae96d07d4a 4 contexts x 63 scored positions = 252 scored positions, score_from 0 sealed: yes (token digest d0ae96d07d4a9888...) -- contamination scan: NOT RUN Reference (teacher) reference--fixture.qwen4-exp.cbec16fe7bff9a88 -- native_bf16, artifact artifact--malaiwah.qwen4-exp-tiny-random-bf16.90fcdd52d45a @90fcdd52d45a200a1b6d446886638af14957117c Metric mean_tokenwise_kld, direction reference_to_candidate, accumulation float64 Estimation surface stack_relation same_stack, head_policy native_head Comparability key cmp--57fd571961121fe6 Like-for-like predicate comparable: unknown -- no recorded difference, but hardware, replay_backend, replay_env, stack are unrecorded for at least one member, so homogeneity cannot be certified. Machine-readable form with per-dimension values: this key's comparability block in index.json.

What this table is. Every row here shares the comparability key above: the same tokens, the same teacher capture, the same metric and direction, the same estimator precision, the same stack relation and the same head policy. That makes them CANDIDATES for ranking -- the key is a necessary partition, not a certificate. Whether they are also like-for-like on the dimensions the key omits (lane, pipeline, scope coverage, hardware) is what the predicate line above answers.

Rank is not a verdict. The table is sorted by fidelity alone, and fidelity buys bits: a larger, higher-bitrate quant will usually sit above a smaller one, which is not news. Read the Size and Codec columns before reading the order, and compare like against like.

What it is NOT comparable to. Every other table in this file: no other group shares this key. That includes every table for a different model -- a KL number is a divergence over one model's own vocabulary against that model's own teacher, never a score that can be carried between models.

Single-row group. This number has nothing in the registry to be ranked against. It is a stated fact, not a placing.

1 of this group's 1 rows came off a different measurement lane (other) and are tabled on their own below, not mixed into the ordering here. The key does not carry the lane; this file does.

Lane other -- 1 of this group's 1 rows

A different lane. Same key, and that is exactly the problem this table solves. The comparability key is a function of the panel, the teacher, the metric, the direction, the estimator precision, the stack relation and the head policy -- and these rows match the table above on all seven. What they do not share is the machine and the code path that produced the candidate logits, and lanes are not interchangeable. Sorting them into one list would read as a ranking; where the same artifact appears in both, it is one set of weights measured twice, not two quants.

No bridge to the sealed lane is recorded for this lane.

Artifact	Codec	Size	mean_tokenwise_kld (nats)	CI95	Top-1	Runs	Attribution	Receipt
qwen4-exp-tiny-random-bf16 (random test fixture) native BF16 (measurement floor)	bf16	0.0 GB	0	--	100.00 %	2 runs, bitwise identical	measured by us	receipt

Bias on qwen4-exp-tiny-random-bf16 (random test fixture) native BF16 -- other, direction unknown. This is the fixture's local CPU reproduction floor. Its offset against a production GPU lane was not measured; no cross-lane equivalence is claimed.

Disclosures for the rows above (3)

• fixture.qwen4-exp.floor.8ba57920f6e4f6ba reduced_run_count: Receipt disclosure reduced_run_count: this dataset is ONE cold capture (run_count 1, the DET-D4 floor is 5). Cross-run determinism is not asserted by the dataset; it is established by a second cold capture plus fidelity-dataset compare --self-compare, whose exactly-0.0 result is the SC-1 reproduction confirmation.
• fixture.qwen4-exp.floor.8ba57920f6e4f6ba non_sealed_lane: Local CPU fixture lane; this exact same-lane reproduction is not a measured offset against a production sealed GPU lane.
• fixture.qwen4-exp.floor.8ba57920f6e4f6ba reduced_run_count: Two qualified cold runs, not five.

SmolLM2-135M native BF16 (unchanged source mirror)

model--malaiwah.smollm2-135m-qfs-native-bf16 -- published by HuggingFaceTB. Tokenizer tokenizer-51666963fa4cef6fbd450fc7, vocabulary 49152.

Panel: SmolLM2-135M native BF16 (unchanged source mirror) qualified token panel

Panel disclosure -- reduced_run_count: reduced_run_count: this dataset is ONE cold capture (run_count 1, the DET-D4 floor is 5). Cross-run determinism is not asserted by the dataset; it is established by a second cold capture plus fidelity-dataset compare --self-compare, whose exactly-0.0 result is the SC-1 reproduction confirmation.

Panel disclosure -- native_head_replay: native_head_replay: HEAD-1d: each side replayed through its own sealed head (reference d4d9c148303f, candidate d4d9c148303f); head-weight differences are included, but replay arithmetic need not equal live/native head execution. The heads are content-identical.

This panel carries 2 separate comparability groups. They are different measurements of different things and are never merged.

Group cmp--4a450f9150f1ec9d -- 3 rows

Panel panel--native.6a74033484e89365fcd6330f -- SmolLM2-135M native BF16 (unchanged source mirror) qualified token panel 16 contexts x 255 scored positions = 4,080 scored positions, score_from 0 sealed: yes (token digest 551bf02836eacae2...) -- contamination scan: NOT RUN Reference (teacher) reference--native.37fc85ce320a80533db7cb94 -- native_bf16, artifact artifact--malaiwah.smollm2-135m-qfs-native-bf16.5dc8abe390e2 @5dc8abe390e2db0102e6afd1509f68fbb5d54e6c Metric mean_tokenwise_kld, direction reference_to_candidate, accumulation float64 Estimation surface stack_relation cross_stack, head_policy native_head Comparability key cmp--4a450f9150f1ec9d Like-for-like predicate comparable: unknown -- no recorded difference, but hardware, stack are unrecorded for at least one member, so homogeneity cannot be certified. Machine-readable form with per-dimension values: this key's comparability block in index.json.

What this table is. Every row here shares the comparability key above: the same tokens, the same teacher capture, the same metric and direction, the same estimator precision, the same stack relation and the same head policy. That makes them CANDIDATES for ranking -- the key is a necessary partition, not a certificate. Whether they are also like-for-like on the dimensions the key omits (lane, pipeline, scope coverage, hardware) is what the predicate line above answers.

Rank is not a verdict. The table is sorted by fidelity alone, and fidelity buys bits: a larger, higher-bitrate quant will usually sit above a smaller one, which is not news. Read the Size and Codec columns before reading the order, and compare like against like.

What it is NOT comparable to. The nearest neighbouring groups differ in:

• cmp--8797e169cd3f4164 (1 row): stack_relation cross_stack -> same_stack

Those numbers are in this file, under their own headings. Quoting one under the other heading is the mistake this layout exists to prevent: the key is a function of the panel, the teacher, the metric, the direction and the estimator, and the validator recomputes it from those fields rather than trusting the stamped value. What that catches is a row filed under a key its own fields do not produce. It does not catch a number attributed to the wrong panel in the first place -- no offline checker can. That is what the receipt digests on every row are for.

Also, and always: every table for a different model. A KL number is a divergence over one model's own vocabulary against that model's own teacher. It is not a quality score that can be carried between models.

3 of this group's 3 rows came off a different measurement lane (other) and are tabled on their own below, not mixed into the ordering here. The key does not carry the lane; this file does.

Lane other -- 3 of this group's 3 rows

A different lane. Same key, and that is exactly the problem this table solves. The comparability key is a function of the panel, the teacher, the metric, the direction, the estimator precision, the stack relation and the head policy -- and these rows match the table above on all seven. What they do not share is the machine and the code path that produced the candidate logits, and lanes are not interchangeable. Sorting them into one list would read as a ranking; where the same artifact appears in both, it is one set of weights measured twice, not two quants.

No bridge to the sealed lane is recorded for this lane.

Artifact	Codec	Size	mean_tokenwise_kld (nats)	CI95	Top-1	Runs	Attribution	Receipt
malaiwah/SmolLM2-135M-QFS-gptq-int4-g32 (measurement floor)	int4 @4	0.1 GB	0.150845	--	81.10 %	1 run, unevidenced	measured by us	receipt
malaiwah/SmolLM2-135M-QFS-gptq-int4-g64 (measurement floor)	int4 @4	0.1 GB	0.178289	--	79.29 %	1 run, unevidenced	measured by us	receipt
malaiwah/SmolLM2-135M-QFS-rtn-int4-g64 (measurement floor)	int4 @4	0.1 GB	0.258016	--	73.24 %	1 run, unevidenced	measured by us	receipt

Bias on malaiwah/SmolLM2-135M-QFS-gptq-int4-g32 -- cross_stack_capture_replay, direction unknown. the capture stacks differ or their code identity is unproven. The sign and magnitude of the cross-stack effect are unknown; a control measured on another artifact, panel or stack is not a transferable correction. usable_as_floor is false.

Disclosures for the rows above (12)

• malaiwah.smollm2-135m-qfs-gptq-int4-g32.native.6a74033484e89365fcd6330f cross_stack_capture: capture stack equality is not established (lane_identity equal on both sides, or unrecorded, stack_fingerprint equal on both sides, or unrecorded, source_files differ or unrecorded). The cross-stack effect has unknown sign and magnitude.
• malaiwah.smollm2-135m-qfs-gptq-int4-g32.native.6a74033484e89365fcd6330f weights_reconstructed: candidate was captured from a WEIGHTS-ONLY RECONSTRUCTION (affine-weight-reconstruction, output bfloat16). The stored weights were decoded before the model forward; native quantized GEMM and serving activation arithmetic are not measured. Decoder provenance is in the sealed runtime receipt. The comparison is advisory.
• malaiwah.smollm2-135m-qfs-gptq-int4-g32.native.6a74033484e89365fcd6330f non_sealed_lane: Produced by the 'other' lane, not the sealed-ep8 lane that the other rows in this comparability group used. Lanes are not interchangeable: this row carries an undisclosed offset against the sealed lane on the same panel until that offset is itself measured and recorded here.
• malaiwah.smollm2-135m-qfs-gptq-int4-g32.native.6a74033484e89365fcd6330f record_note: This service read authenticated HF provider metadata and validated recovered evidence; it did not independently reproduce the model run.
• malaiwah.smollm2-135m-qfs-gptq-int4-g64.native.6a74033484e89365fcd6330f cross_stack_capture: capture stack equality is not established (lane_identity equal on both sides, or unrecorded, stack_fingerprint equal on both sides, or unrecorded, source_files differ or unrecorded). The cross-stack effect has unknown sign and magnitude.
• malaiwah.smollm2-135m-qfs-gptq-int4-g64.native.6a74033484e89365fcd6330f weights_reconstructed: candidate was captured from a WEIGHTS-ONLY RECONSTRUCTION (affine-weight-reconstruction, output bfloat16). The stored weights were decoded before the model forward; native quantized GEMM and serving activation arithmetic are not measured. Decoder provenance is in the sealed runtime receipt. The comparison is advisory.
• malaiwah.smollm2-135m-qfs-gptq-int4-g64.native.6a74033484e89365fcd6330f non_sealed_lane: Produced by the 'other' lane, not the sealed-ep8 lane that the other rows in this comparability group used. Lanes are not interchangeable: this row carries an undisclosed offset against the sealed lane on the same panel until that offset is itself measured and recorded here.
• malaiwah.smollm2-135m-qfs-gptq-int4-g64.native.6a74033484e89365fcd6330f record_note: This service read authenticated HF provider metadata and validated recovered evidence; it did not independently reproduce the model run.
• malaiwah.smollm2-135m-qfs-rtn-int4-g64.native.6a74033484e89365fcd6330f cross_stack_capture: capture stack equality is not established (lane_identity equal on both sides, or unrecorded, stack_fingerprint equal on both sides, or unrecorded, source_files differ or unrecorded). The cross-stack effect has unknown sign and magnitude.
• malaiwah.smollm2-135m-qfs-rtn-int4-g64.native.6a74033484e89365fcd6330f weights_reconstructed: candidate was captured from a WEIGHTS-ONLY RECONSTRUCTION (affine-weight-reconstruction, output bfloat16). The stored weights were decoded before the model forward; native quantized GEMM and serving activation arithmetic are not measured. Decoder provenance is in the sealed runtime receipt. The comparison is advisory.
• malaiwah.smollm2-135m-qfs-rtn-int4-g64.native.6a74033484e89365fcd6330f non_sealed_lane: Produced by the 'other' lane, not the sealed-ep8 lane that the other rows in this comparability group used. Lanes are not interchangeable: this row carries an undisclosed offset against the sealed lane on the same panel until that offset is itself measured and recorded here.
• malaiwah.smollm2-135m-qfs-rtn-int4-g64.native.6a74033484e89365fcd6330f record_note: This service read authenticated HF provider metadata and validated recovered evidence; it did not independently reproduce the model run.

Group cmp--8797e169cd3f4164 -- 1 row

Panel panel--native.6a74033484e89365fcd6330f -- SmolLM2-135M native BF16 (unchanged source mirror) qualified token panel 16 contexts x 255 scored positions = 4,080 scored positions, score_from 0 sealed: yes (token digest 551bf02836eacae2...) -- contamination scan: NOT RUN Reference (teacher) reference--native.37fc85ce320a80533db7cb94 -- native_bf16, artifact artifact--malaiwah.smollm2-135m-qfs-native-bf16.5dc8abe390e2 @5dc8abe390e2db0102e6afd1509f68fbb5d54e6c Metric mean_tokenwise_kld, direction reference_to_candidate, accumulation float64 Estimation surface stack_relation same_stack, head_policy native_head Comparability key cmp--8797e169cd3f4164 Like-for-like predicate comparable: unknown -- no recorded difference, but hardware, replay_backend, replay_env, stack are unrecorded for at least one member, so homogeneity cannot be certified. Machine-readable form with per-dimension values: this key's comparability block in index.json.

What this table is. Every row here shares the comparability key above: the same tokens, the same teacher capture, the same metric and direction, the same estimator precision, the same stack relation and the same head policy. That makes them CANDIDATES for ranking -- the key is a necessary partition, not a certificate. Whether they are also like-for-like on the dimensions the key omits (lane, pipeline, scope coverage, hardware) is what the predicate line above answers.

Rank is not a verdict. The table is sorted by fidelity alone, and fidelity buys bits: a larger, higher-bitrate quant will usually sit above a smaller one, which is not news. Read the Size and Codec columns before reading the order, and compare like against like.

What it is NOT comparable to. The nearest neighbouring groups differ in:

• cmp--4a450f9150f1ec9d (3 rows): stack_relation same_stack -> cross_stack

Those numbers are in this file, under their own headings. Quoting one under the other heading is the mistake this layout exists to prevent: the key is a function of the panel, the teacher, the metric, the direction and the estimator, and the validator recomputes it from those fields rather than trusting the stamped value. What that catches is a row filed under a key its own fields do not produce. It does not catch a number attributed to the wrong panel in the first place -- no offline checker can. That is what the receipt digests on every row are for.

Also, and always: every table for a different model. A KL number is a divergence over one model's own vocabulary against that model's own teacher. It is not a quality score that can be carried between models.

Single-row group. This number has nothing in the registry to be ranked against. It is a stated fact, not a placing.

1 of this group's 1 rows came off a different measurement lane (other) and are tabled on their own below, not mixed into the ordering here. The key does not carry the lane; this file does.

Lane other -- 1 of this group's 1 rows

A different lane. Same key, and that is exactly the problem this table solves. The comparability key is a function of the panel, the teacher, the metric, the direction, the estimator precision, the stack relation and the head policy -- and these rows match the table above on all seven. What they do not share is the machine and the code path that produced the candidate logits, and lanes are not interchangeable. Sorting them into one list would read as a ranking; where the same artifact appears in both, it is one set of weights measured twice, not two quants.

No bridge to the sealed lane is recorded for this lane.

Artifact	Codec	Size	mean_tokenwise_kld (nats)	CI95	Top-1	Runs	Attribution	Receipt
SmolLM2-135M native BF16 (unchanged source mirror) native BF16 (measurement floor)	bf16	0.3 GB	0	--	100.00 %	2 runs, bitwise identical	measured by us	receipt

Bias on SmolLM2-135M native BF16 (unchanged source mirror) native BF16 -- other, direction unknown. Exact same-lane native reproduction; its offset against other lanes was not measured.

Disclosures for the rows above (4)

• native.floor.b3814ec840800c44fb3f531b reduced_run_count: reduced_run_count: this dataset is ONE cold capture (run_count 1, the DET-D4 floor is 5). Cross-run determinism is not asserted by the dataset; it is established by a second cold capture plus fidelity-dataset compare --self-compare, whose exactly-0.0 result is the SC-1 reproduction confirmation.
• native.floor.b3814ec840800c44fb3f531b reduced_run_count: Two qualified cold runs, not five.
• native.floor.b3814ec840800c44fb3f531b non_sealed_lane: This native reproduction is confined to its recorded HF Jobs lane; no production-lane equivalence is established.
• native.floor.b3814ec840800c44fb3f531b record_note: This service read authenticated HF provider metadata and validated recovered evidence; it did not independently reproduce the model run.

spark2-5-tiny-random-bf16 (random test fixture)

model--malaiwah.spark2-5-tiny-random-bf16 -- published by malaiwah. Tokenizer fixture-tokenizer-08d3f8f92d3b9086, vocabulary 260.

Panel: Synthetic CPU fixture panel 525cb6c62509

Panel disclosure -- reduced_run_count: Receipt disclosure reduced_run_count: this dataset is ONE cold capture (run_count 1, the DET-D4 floor is 5). Cross-run determinism is not asserted by the dataset; it is established by a second cold capture plus fidelity-dataset compare --self-compare, whose exactly-0.0 result is the SC-1 reproduction confirmation.

Panel disclosure -- native_head_replay: Receipt disclosure native_head_replay: HEAD-1d: each side replayed through its own sealed head (reference 2364995d4009, candidate 2364995d4009); head error is inside the measurement, as under HEAD-2, and nothing is substituted. The heads are content-identical.

Group cmp--1b9cc52aa250300f -- 1 row

Panel panel--fixture.9d187d1dab31575b41ab4b00 -- Synthetic CPU fixture panel 525cb6c62509 4 contexts x 63 scored positions = 252 scored positions, score_from 0 sealed: yes (token digest 525cb6c625096ddb...) -- contamination scan: NOT RUN Reference (teacher) reference--fixture.spark2-5.022a11db4e1e3331 -- native_bf16, artifact artifact--malaiwah.spark2-5-tiny-random-bf16.c0a57b037007 @c0a57b037007d475ede719e78f4a6ce24f211da0 Metric mean_tokenwise_kld, direction reference_to_candidate, accumulation float64 Estimation surface stack_relation same_stack, head_policy native_head Comparability key cmp--1b9cc52aa250300f Like-for-like predicate comparable: unknown -- no recorded difference, but hardware, replay_backend, replay_env, stack are unrecorded for at least one member, so homogeneity cannot be certified. Machine-readable form with per-dimension values: this key's comparability block in index.json.

What this table is. Every row here shares the comparability key above: the same tokens, the same teacher capture, the same metric and direction, the same estimator precision, the same stack relation and the same head policy. That makes them CANDIDATES for ranking -- the key is a necessary partition, not a certificate. Whether they are also like-for-like on the dimensions the key omits (lane, pipeline, scope coverage, hardware) is what the predicate line above answers.

Rank is not a verdict. The table is sorted by fidelity alone, and fidelity buys bits: a larger, higher-bitrate quant will usually sit above a smaller one, which is not news. Read the Size and Codec columns before reading the order, and compare like against like.

What it is NOT comparable to. Every other table in this file: no other group shares this key. That includes every table for a different model -- a KL number is a divergence over one model's own vocabulary against that model's own teacher, never a score that can be carried between models.

Single-row group. This number has nothing in the registry to be ranked against. It is a stated fact, not a placing.

1 of this group's 1 rows came off a different measurement lane (other) and are tabled on their own below, not mixed into the ordering here. The key does not carry the lane; this file does.

Lane other -- 1 of this group's 1 rows

A different lane. Same key, and that is exactly the problem this table solves. The comparability key is a function of the panel, the teacher, the metric, the direction, the estimator precision, the stack relation and the head policy -- and these rows match the table above on all seven. What they do not share is the machine and the code path that produced the candidate logits, and lanes are not interchangeable. Sorting them into one list would read as a ranking; where the same artifact appears in both, it is one set of weights measured twice, not two quants.

No bridge to the sealed lane is recorded for this lane.

Artifact	Codec	Size	mean_tokenwise_kld (nats)	CI95	Top-1	Runs	Attribution	Receipt
spark2-5-tiny-random-bf16 (random test fixture) native BF16 (measurement floor)	bf16	0.0 GB	0	--	100.00 %	2 runs, bitwise identical	measured by us	receipt

Bias on spark2-5-tiny-random-bf16 (random test fixture) native BF16 -- other, direction unknown. This is the fixture's local CPU reproduction floor. Its offset against a production GPU lane was not measured; no cross-lane equivalence is claimed.

Disclosures for the rows above (3)

• fixture.spark2-5.floor.d62ef495a49ba3a1 reduced_run_count: Receipt disclosure reduced_run_count: this dataset is ONE cold capture (run_count 1, the DET-D4 floor is 5). Cross-run determinism is not asserted by the dataset; it is established by a second cold capture plus fidelity-dataset compare --self-compare, whose exactly-0.0 result is the SC-1 reproduction confirmation.
• fixture.spark2-5.floor.d62ef495a49ba3a1 non_sealed_lane: Local CPU fixture lane; this exact same-lane reproduction is not a measured offset against a production sealed GPU lane.
• fixture.spark2-5.floor.d62ef495a49ba3a1 reduced_run_count: Two qualified cold runs, not five.

Qwen3.8-27B

model--qwen.qwen3.8-27b -- published by Qwen (Alibaba). Tokenizer qwen3.8, vocabulary 248320.

Panel: malaiwah Qwen3.8-27B distribution-fidelity suite v5 -- 5,120 contexts

Panel disclosure -- unsealed_source: The qwen38 v5 token suite is pinned by suite_token_sha256 and by its manifest digest c79dfad3..., but the token files themselves are not published, so a third party cannot reproduce the digest today.

Group cmp--c8c4df32774bdb63 -- 6 rows

Panel panel--qwen38.malaiwah.suite-v5-10m -- malaiwah Qwen3.8-27B distribution-fidelity suite v5 -- 5,120 contexts 5120 contexts x 2047 scored positions = 10,480,640 scored positions, score_from 0 sealed: yes (token digest 510541f6861b589d...) -- contamination scan: yes, 0 hits Reference (teacher) reference--malaiwah.qwen38-bf16-vllm.suite-v5-10m -- native_bf16, artifact artifact--qwen.qwen3.8-27b-bf16 @unpinned revision Metric mean_tokenwise_kld, direction reference_to_candidate, accumulation float32_reduce_legacy Estimation surface stack_relation same_stack, head_policy shared_reference_head Comparability key cmp--c8c4df32774bdb63 Like-for-like predicate comparable: false -- a RECORDED secondary dimension differs across members: scope. Equal keys make these rows candidates for comparison, not certified like-for-like; ranking across the differing dimension attributes a lane/pipeline/hardware/scope effect to quantization quality. Machine-readable form with per-dimension values: this key's comparability block in index.json.

What this table is. Every row here shares the comparability key above: the same tokens, the same teacher capture, the same metric and direction, the same estimator precision, the same stack relation and the same head policy. That makes them CANDIDATES for ranking -- the key is a necessary partition, not a certificate. Whether they are also like-for-like on the dimensions the key omits (lane, pipeline, scope coverage, hardware) is what the predicate line above answers.

Rank is not a verdict. The table is sorted by fidelity alone, and fidelity buys bits: a larger, higher-bitrate quant will usually sit above a smaller one, which is not news. Read the Size and Codec columns before reading the order, and compare like against like.

What it is NOT comparable to. The nearest neighbouring groups differ in:

• cmp--75b64be1f101ed22 (12 rows): panel_id panel--qwen38.malaiwah.suite-v5-10m -> panel--qwen38.malaiwah.suite-v5-shard0-1m; reference_id reference--malaiwah.qwen38-bf16-vllm.suite-v5-10m -> reference--malaiwah.qwen38-bf16-vllm.suite-v5-shard0-1m
• cmp--726ac1b18b8129fa (5 rows): panel_id panel--qwen38.malaiwah.suite-v5-10m -> panel--qwen38.malaiwah.suite-v5-shard0-1m.scorefrom1024; reference_id reference--malaiwah.qwen38-bf16-vllm.suite-v5-10m -> reference--malaiwah.qwen38-bf16-vllm.suite-v5-shard0-1m.scorefrom1024
• cmp--0bb49e8411b6dc75 (5 rows): panel_id panel--qwen38.malaiwah.suite-v5-10m -> panel--qwen38.malaiwah.suite-v5-shard0-1m.scorefrom256; reference_id reference--malaiwah.qwen38-bf16-vllm.suite-v5-10m -> reference--malaiwah.qwen38-bf16-vllm.suite-v5-shard0-1m.scorefrom256
• cmp--47c0bc74ebec3fa7 (5 rows): panel_id panel--qwen38.malaiwah.suite-v5-10m -> panel--qwen38.malaiwah.suite-v5-shards01-2m; reference_id reference--malaiwah.qwen38-bf16-vllm.suite-v5-10m -> reference--malaiwah.qwen38-bf16-vllm.suite-v5-shards01-2m

Those numbers are in this file, under their own headings. Quoting one under the other heading is the mistake this layout exists to prevent: the key is a function of the panel, the teacher, the metric, the direction and the estimator, and the validator recomputes it from those fields rather than trusting the stamped value. What that catches is a row filed under a key its own fields do not produce. It does not catch a number attributed to the wrong panel in the first place -- no offline checker can. That is what the receipt digests on every row are for.

Also, and always: every table for a different model. A KL number is a divergence over one model's own vocabulary against that model's own teacher. It is not a quality score that can be carried between models.

Artifact	Codec	Size	mean_tokenwise_kld (nats)	CI95	Top-1	Runs	Attribution	Receipt
malaiwah Qwen3.8-27B EXL3 K5K6 hydrated	exl3-mcg @5	21.6 GB	0.00275963	[0.00254024, 0.00302032]	97.70 %	1 run, unevidenced	measured by us	receipt
malaiwah Qwen3.8-27B EXL3 K5K6	exl3-mcg @5	30.6 GB	0.00320988	[0.00298238, 0.00348017]	97.52 %	1 run, unevidenced	measured by us	receipt
malaiwah Qwen3.8-27B EXL3 K5K6 context	exl3-mcg @5	20.7 GB	0.00350936	[0.00321967, 0.00385239]	97.44 %	1 run, unevidenced	measured by us	receipt
Qwen3.8-27B FP8 (official)	fp8_e4m3 @8	30.9 GB	0.00529394	[0.00492736, 0.00572785]	96.79 %	1 run, unevidenced	measured by us (their artifact)	receipt
malaiwah Qwen3.8-27B K4	exl3-mcg @4	28.3 GB	0.0106039	[0.00963981, 0.0117463]	95.76 %	1 run, unevidenced	measured by us	receipt
unsloth Qwen3.8-27B NVFP4	nvfp4 @4	--	0.0310586	[0.0279161, 0.0347947]	92.90 %	1 run, unevidenced	measured by us (their artifact)	receipt

The same artifact, measured elsewhere in this file. 6 of the artifacts below also carry a number in another table -- on a different panel, teacher or estimator -- and the widest of those spans 77%. None of the readings is wrong and none is interchangeable with another. Quoting one of them as the number for the artifact, without its table, is the misuse this registry exists to make obvious.

the 6 artifacts and their ranges

• Qwen3.8-27B FP8 (official) -- 6 values here, from 0.00298985 to 0.00529563 nats (77% apart). Other tables: cmp--05e16411a5932713, cmp--0bb49e8411b6dc75, cmp--47c0bc74ebec3fa7, cmp--726ac1b18b8129fa, cmp--75b64be1f101ed22.
• malaiwah Qwen3.8-27B EXL3 K5K6 -- 5 values here, from 0.0030196 to 0.00320988 nats (6% apart). Other tables: cmp--0bb49e8411b6dc75, cmp--47c0bc74ebec3fa7, cmp--726ac1b18b8129fa, cmp--75b64be1f101ed22.
• malaiwah Qwen3.8-27B EXL3 K5K6 context -- 5 values here, from 0.00324322 to 0.00350936 nats (8% apart). Other tables: cmp--0bb49e8411b6dc75, cmp--47c0bc74ebec3fa7, cmp--726ac1b18b8129fa, cmp--75b64be1f101ed22.
• malaiwah Qwen3.8-27B EXL3 K5K6 hydrated -- 5 values here, from 0.00257964 to 0.00275963 nats (7% apart). Other tables: cmp--0bb49e8411b6dc75, cmp--47c0bc74ebec3fa7, cmp--726ac1b18b8129fa, cmp--75b64be1f101ed22.
• malaiwah Qwen3.8-27B K4 -- 5 values here, from 0.00987561 to 0.0106039 nats (7% apart). Other tables: cmp--0bb49e8411b6dc75, cmp--47c0bc74ebec3fa7, cmp--726ac1b18b8129fa, cmp--75b64be1f101ed22.
• unsloth Qwen3.8-27B NVFP4 -- 2 values here, from 0.0301154 to 0.0310586 nats (3% apart). Other tables: cmp--75b64be1f101ed22.

Disclosures for the rows above (19)

• qwen38.k5k6-hydrated.suite-v5-10m revision_unpinned: No measurement receipt for this artifact records a Hub revision. Every kld5 receipt records model_revision=null / model_revision_source='none'. Identity rests on index_sha256 and the per-shard sha256 map the receipt carries.
• qwen38.k5k6-hydrated.suite-v5-10m single_run: One pass. Repeatability was not established for this row.
• qwen38.k5k6-hydrated.suite-v5-10m fp32_vocab_reduction: ESTIMATOR DEFECT, disclosed 2026-08-31 (P1-06). The scorer computed the vocabulary reduction in float32 and cast the finished sum to float64; this row previously declared accumulation_dtype float64. Relabeled float32_reduce_legacy -- the value is unchanged, the comparability key moved, and the row ranks only against rows from the same float32-reducing scorer. Synthetic worst case for the defect class: negative per-token 'KL' near -1e-6 against a true value of ~2e-8 on near-equal distributions; this ladder's published means sit at 1e-3..1e-1, three to five orders above that error scale. See docs/PUBLISHED-CORRECTIONS.md.
• qwen38.k5k6.suite-v5-10m revision_unpinned: No measurement receipt for this artifact records a Hub revision. Every kld5 receipt records model_revision=null / model_revision_source='none'. Identity rests on index_sha256 and the per-shard sha256 map the receipt carries.
• qwen38.k5k6.suite-v5-10m single_run: One pass. Repeatability was not established for this row.
• qwen38.k5k6.suite-v5-10m fp32_vocab_reduction: ESTIMATOR DEFECT, disclosed 2026-08-31 (P1-06). The scorer computed the vocabulary reduction in float32 and cast the finished sum to float64; this row previously declared accumulation_dtype float64. Relabeled float32_reduce_legacy -- the value is unchanged, the comparability key moved, and the row ranks only against rows from the same float32-reducing scorer. Synthetic worst case for the defect class: negative per-token 'KL' near -1e-6 against a true value of ~2e-8 on near-equal distributions; this ladder's published means sit at 1e-3..1e-1, three to five orders above that error scale. See docs/PUBLISHED-CORRECTIONS.md.
• qwen38.k5k6-context.suite-v5-10m revision_unpinned: No measurement receipt for this artifact records a Hub revision. Every kld5 receipt records model_revision=null / model_revision_source='none'. Identity rests on index_sha256 and the per-shard sha256 map the receipt carries.
• qwen38.k5k6-context.suite-v5-10m single_run: One pass. Repeatability was not established for this row.
• qwen38.k5k6-context.suite-v5-10m fp32_vocab_reduction: ESTIMATOR DEFECT, disclosed 2026-08-31 (P1-06). The scorer computed the vocabulary reduction in float32 and cast the finished sum to float64; this row previously declared accumulation_dtype float64. Relabeled float32_reduce_legacy -- the value is unchanged, the comparability key moved, and the row ranks only against rows from the same float32-reducing scorer. Synthetic worst case for the defect class: negative per-token 'KL' near -1e-6 against a true value of ~2e-8 on near-equal distributions; this ladder's published means sit at 1e-3..1e-1, three to five orders above that error scale. See docs/PUBLISHED-CORRECTIONS.md.
• qwen38.official-fp8.suite-v5-10m revision_unpinned: No measurement receipt for this artifact records a Hub revision. Every kld5 receipt records model_revision=null / model_revision_source='none'. Identity rests on index_sha256 and the per-shard sha256 map the receipt carries.
• qwen38.official-fp8.suite-v5-10m single_run: One pass. Repeatability was not established for this row.
• qwen38.official-fp8.suite-v5-10m fp32_vocab_reduction: ESTIMATOR DEFECT, disclosed 2026-08-31 (P1-06). The scorer computed the vocabulary reduction in float32 and cast the finished sum to float64; this row previously declared accumulation_dtype float64. Relabeled float32_reduce_legacy -- the value is unchanged, the comparability key moved, and the row ranks only against rows from the same float32-reducing scorer. Synthetic worst case for the defect class: negative per-token 'KL' near -1e-6 against a true value of ~2e-8 on near-equal distributions; this ladder's published means sit at 1e-3..1e-1, three to five orders above that error scale. See docs/PUBLISHED-CORRECTIONS.md.
• qwen38.k4.suite-v5-10m revision_unpinned: No measurement receipt for this artifact records a Hub revision. Every kld5 receipt records model_revision=null / model_revision_source='none'. Identity rests on index_sha256 and the per-shard sha256 map the receipt carries.
• qwen38.k4.suite-v5-10m single_run: One pass. Repeatability was not established for this row.
• qwen38.k4.suite-v5-10m fp32_vocab_reduction: ESTIMATOR DEFECT, disclosed 2026-08-31 (P1-06). The scorer computed the vocabulary reduction in float32 and cast the finished sum to float64; this row previously declared accumulation_dtype float64. Relabeled float32_reduce_legacy -- the value is unchanged, the comparability key moved, and the row ranks only against rows from the same float32-reducing scorer. Synthetic worst case for the defect class: negative per-token 'KL' near -1e-6 against a true value of ~2e-8 on near-equal distributions; this ladder's published means sit at 1e-3..1e-1, three to five orders above that error scale. See docs/PUBLISHED-CORRECTIONS.md.
• qwen38.unsloth-nvfp4.suite-v5-10m revision_unpinned: No measurement receipt for this artifact records a Hub revision. Every kld5 receipt records model_revision=null / model_revision_source='none'. Identity rests on index_sha256 and the per-shard sha256 map the receipt carries.
• qwen38.unsloth-nvfp4.suite-v5-10m artifact_identity_incomplete: The per-tensor-class quantization recipe for this artifact was never published, so scope.assignments records 'unknown' rather than a guessed allocation. Its scope_digest shows the gap.
• qwen38.unsloth-nvfp4.suite-v5-10m single_run: One pass. Repeatability was not established for this row.
• qwen38.unsloth-nvfp4.suite-v5-10m fp32_vocab_reduction: ESTIMATOR DEFECT, disclosed 2026-08-31 (P1-06). The scorer computed the vocabulary reduction in float32 and cast the finished sum to float64; this row previously declared accumulation_dtype float64. Relabeled float32_reduce_legacy -- the value is unchanged, the comparability key moved, and the row ranks only against rows from the same float32-reducing scorer. Synthetic worst case for the defect class: negative per-token 'KL' near -1e-6 against a true value of ~2e-8 on near-equal distributions; this ladder's published means sit at 1e-3..1e-1, three to five orders above that error scale. See docs/PUBLISHED-CORRECTIONS.md.

Panel: malaiwah Qwen3.8-27B suite v5, shards 0-1 -- 1,024 contexts

Derived from panel--qwen38.malaiwah.suite-v5-10m by shard_subset: shards 0 and 1 of 10 (1,024 contexts, 495 source clusters).

Panel disclosure -- unsealed_source: No combined token digest was published for the two-shard union; the two per-shard digests are recorded instead, which pin the content but are not a single panel identity.

Panel disclosure -- unsealed_source: The qwen38 v5 token suite is pinned by suite_token_sha256 and by its manifest digest c79dfad3..., but the token files themselves are not published, so a third party cannot reproduce the digest today.

Group cmp--47c0bc74ebec3fa7 -- 5 rows

Panel panel--qwen38.malaiwah.suite-v5-shards01-2m -- malaiwah Qwen3.8-27B suite v5, shards 0-1 -- 1,024 contexts 1024 contexts x 2047 scored positions = 2,096,128 scored positions, score_from 0 sealed: no -- contamination scan: yes, 0 hits Reference (teacher) reference--malaiwah.qwen38-bf16-vllm.suite-v5-shards01-2m -- native_bf16, artifact artifact--qwen.qwen3.8-27b-bf16 @unpinned revision Metric mean_tokenwise_kld, direction reference_to_candidate, accumulation float32_reduce_legacy Estimation surface stack_relation same_stack, head_policy shared_reference_head Comparability key cmp--47c0bc74ebec3fa7 Like-for-like predicate comparable: false -- a RECORDED secondary dimension differs across members: scope. Equal keys make these rows candidates for comparison, not certified like-for-like; ranking across the differing dimension attributes a lane/pipeline/hardware/scope effect to quantization quality. Machine-readable form with per-dimension values: this key's comparability block in index.json.

What this table is. Every row here shares the comparability key above: the same tokens, the same teacher capture, the same metric and direction, the same estimator precision, the same stack relation and the same head policy. That makes them CANDIDATES for ranking -- the key is a necessary partition, not a certificate. Whether they are also like-for-like on the dimensions the key omits (lane, pipeline, scope coverage, hardware) is what the predicate line above answers.

Rank is not a verdict. The table is sorted by fidelity alone, and fidelity buys bits: a larger, higher-bitrate quant will usually sit above a smaller one, which is not news. Read the Size and Codec columns before reading the order, and compare like against like.

What it is NOT comparable to. The nearest neighbouring groups differ in:

• cmp--75b64be1f101ed22 (12 rows): panel_id panel--qwen38.malaiwah.suite-v5-shards01-2m -> panel--qwen38.malaiwah.suite-v5-shard0-1m; reference_id reference--malaiwah.qwen38-bf16-vllm.suite-v5-shards01-2m -> reference--malaiwah.qwen38-bf16-vllm.suite-v5-shard0-1m
• cmp--c8c4df32774bdb63 (6 rows): panel_id panel--qwen38.malaiwah.suite-v5-shards01-2m -> panel--qwen38.malaiwah.suite-v5-10m; reference_id reference--malaiwah.qwen38-bf16-vllm.suite-v5-shards01-2m -> reference--malaiwah.qwen38-bf16-vllm.suite-v5-10m
• cmp--726ac1b18b8129fa (5 rows): panel_id panel--qwen38.malaiwah.suite-v5-shards01-2m -> panel--qwen38.malaiwah.suite-v5-shard0-1m.scorefrom1024; reference_id reference--malaiwah.qwen38-bf16-vllm.suite-v5-shards01-2m -> reference--malaiwah.qwen38-bf16-vllm.suite-v5-shard0-1m.scorefrom1024
• cmp--0bb49e8411b6dc75 (5 rows): panel_id panel--qwen38.malaiwah.suite-v5-shards01-2m -> panel--qwen38.malaiwah.suite-v5-shard0-1m.scorefrom256; reference_id reference--malaiwah.qwen38-bf16-vllm.suite-v5-shards01-2m -> reference--malaiwah.qwen38-bf16-vllm.suite-v5-shard0-1m.scorefrom256

Those numbers are in this file, under their own headings. Quoting one under the other heading is the mistake this layout exists to prevent: the key is a function of the panel, the teacher, the metric, the direction and the estimator, and the validator recomputes it from those fields rather than trusting the stamped value. What that catches is a row filed under a key its own fields do not produce. It does not catch a number attributed to the wrong panel in the first place -- no offline checker can. That is what the receipt digests on every row are for.

Also, and always: every table for a different model. A KL number is a divergence over one model's own vocabulary against that model's own teacher. It is not a quality score that can be carried between models.

Artifact	Codec	Size	mean_tokenwise_kld (nats)	CI95	Top-1	Runs	Attribution	Receipt
malaiwah Qwen3.8-27B EXL3 K5K6 hydrated	exl3-mcg @5	21.6 GB	0.00275854	[0.0025346, 0.00302484]	97.75 %	1 run, unevidenced	measured by us	receipt
malaiwah Qwen3.8-27B EXL3 K5K6	exl3-mcg @5	30.6 GB	0.00320026	[0.00296909, 0.00347268]	97.56 %	1 run, unevidenced	measured by us	receipt
malaiwah Qwen3.8-27B EXL3 K5K6 context	exl3-mcg @5	20.7 GB	0.00350243	[0.00321244, 0.00384432]	97.49 %	1 run, unevidenced	measured by us	receipt
Qwen3.8-27B FP8 (official)	fp8_e4m3 @8	30.9 GB	0.00529563	[0.00492561, 0.00572892]	96.85 %	1 run, unevidenced	measured by us (their artifact)	receipt
malaiwah Qwen3.8-27B K4	exl3-mcg @4	28.3 GB	0.0105726	[0.0096214, 0.0117039]	95.83 %	1 run, unevidenced	measured by us	receipt

The same artifact, measured elsewhere in this file. 5 of the artifacts below also carry a number in another table -- on a different panel, teacher or estimator -- and the widest of those spans 77%. None of the readings is wrong and none is interchangeable with another. Quoting one of them as the number for the artifact, without its table, is the misuse this registry exists to make obvious.

the 5 artifacts and their ranges

• Qwen3.8-27B FP8 (official) -- 6 values here, from 0.00298985 to 0.00529563 nats (77% apart). Other tables: cmp--05e16411a5932713, cmp--0bb49e8411b6dc75, cmp--726ac1b18b8129fa, cmp--75b64be1f101ed22, cmp--c8c4df32774bdb63.
• malaiwah Qwen3.8-27B EXL3 K5K6 -- 5 values here, from 0.0030196 to 0.00320988 nats (6% apart). Other tables: cmp--0bb49e8411b6dc75, cmp--726ac1b18b8129fa, cmp--75b64be1f101ed22, cmp--c8c4df32774bdb63.
• malaiwah Qwen3.8-27B EXL3 K5K6 context -- 5 values here, from 0.00324322 to 0.00350936 nats (8% apart). Other tables: cmp--0bb49e8411b6dc75, cmp--726ac1b18b8129fa, cmp--75b64be1f101ed22, cmp--c8c4df32774bdb63.
• malaiwah Qwen3.8-27B EXL3 K5K6 hydrated -- 5 values here, from 0.00257964 to 0.00275963 nats (7% apart). Other tables: cmp--0bb49e8411b6dc75, cmp--726ac1b18b8129fa, cmp--75b64be1f101ed22, cmp--c8c4df32774bdb63.
• malaiwah Qwen3.8-27B K4 -- 5 values here, from 0.00987561 to 0.0106039 nats (7% apart). Other tables: cmp--0bb49e8411b6dc75, cmp--726ac1b18b8129fa, cmp--75b64be1f101ed22, cmp--c8c4df32774bdb63.

Disclosures for the rows above (15)

• qwen38.k5k6-hydrated.suite-v5-shards01-2m revision_unpinned: No measurement receipt for this artifact records a Hub revision. Every kld5 receipt records model_revision=null / model_revision_source='none'. Identity rests on index_sha256 and the per-shard sha256 map the receipt carries.
• qwen38.k5k6-hydrated.suite-v5-shards01-2m single_run: One pass. Repeatability was not established for this row.
• qwen38.k5k6-hydrated.suite-v5-shards01-2m fp32_vocab_reduction: ESTIMATOR DEFECT, disclosed 2026-08-31 (P1-06). The scorer computed the vocabulary reduction in float32 and cast the finished sum to float64; this row previously declared accumulation_dtype float64. Relabeled float32_reduce_legacy -- the value is unchanged, the comparability key moved, and the row ranks only against rows from the same float32-reducing scorer. Synthetic worst case for the defect class: negative per-token 'KL' near -1e-6 against a true value of ~2e-8 on near-equal distributions; this ladder's published means sit at 1e-3..1e-1, three to five orders above that error scale. See docs/PUBLISHED-CORRECTIONS.md.
• qwen38.k5k6.suite-v5-shards01-2m revision_unpinned: No measurement receipt for this artifact records a Hub revision. Every kld5 receipt records model_revision=null / model_revision_source='none'. Identity rests on index_sha256 and the per-shard sha256 map the receipt carries.
• qwen38.k5k6.suite-v5-shards01-2m single_run: One pass. Repeatability was not established for this row.
• qwen38.k5k6.suite-v5-shards01-2m fp32_vocab_reduction: ESTIMATOR DEFECT, disclosed 2026-08-31 (P1-06). The scorer computed the vocabulary reduction in float32 and cast the finished sum to float64; this row previously declared accumulation_dtype float64. Relabeled float32_reduce_legacy -- the value is unchanged, the comparability key moved, and the row ranks only against rows from the same float32-reducing scorer. Synthetic worst case for the defect class: negative per-token 'KL' near -1e-6 against a true value of ~2e-8 on near-equal distributions; this ladder's published means sit at 1e-3..1e-1, three to five orders above that error scale. See docs/PUBLISHED-CORRECTIONS.md.
• qwen38.k5k6-context.suite-v5-shards01-2m revision_unpinned: No measurement receipt for this artifact records a Hub revision. Every kld5 receipt records model_revision=null / model_revision_source='none'. Identity rests on index_sha256 and the per-shard sha256 map the receipt carries.
• qwen38.k5k6-context.suite-v5-shards01-2m single_run: One pass. Repeatability was not established for this row.
• qwen38.k5k6-context.suite-v5-shards01-2m fp32_vocab_reduction: ESTIMATOR DEFECT, disclosed 2026-08-31 (P1-06). The scorer computed the vocabulary reduction in float32 and cast the finished sum to float64; this row previously declared accumulation_dtype float64. Relabeled float32_reduce_legacy -- the value is unchanged, the comparability key moved, and the row ranks only against rows from the same float32-reducing scorer. Synthetic worst case for the defect class: negative per-token 'KL' near -1e-6 against a true value of ~2e-8 on near-equal distributions; this ladder's published means sit at 1e-3..1e-1, three to five orders above that error scale. See docs/PUBLISHED-CORRECTIONS.md.
• qwen38.official-fp8.suite-v5-shards01-2m revision_unpinned: No measurement receipt for this artifact records a Hub revision. Every kld5 receipt records model_revision=null / model_revision_source='none'. Identity rests on index_sha256 and the per-shard sha256 map the receipt carries.
• qwen38.official-fp8.suite-v5-shards01-2m single_run: One pass. Repeatability was not established for this row.
• qwen38.official-fp8.suite-v5-shards01-2m fp32_vocab_reduction: ESTIMATOR DEFECT, disclosed 2026-08-31 (P1-06). The scorer computed the vocabulary reduction in float32 and cast the finished sum to float64; this row previously declared accumulation_dtype float64. Relabeled float32_reduce_legacy -- the value is unchanged, the comparability key moved, and the row ranks only against rows from the same float32-reducing scorer. Synthetic worst case for the defect class: negative per-token 'KL' near -1e-6 against a true value of ~2e-8 on near-equal distributions; this ladder's published means sit at 1e-3..1e-1, three to five orders above that error scale. See docs/PUBLISHED-CORRECTIONS.md.
• qwen38.k4.suite-v5-shards01-2m revision_unpinned: No measurement receipt for this artifact records a Hub revision. Every kld5 receipt records model_revision=null / model_revision_source='none'. Identity rests on index_sha256 and the per-shard sha256 map the receipt carries.
• qwen38.k4.suite-v5-shards01-2m single_run: One pass. Repeatability was not established for this row.
• qwen38.k4.suite-v5-shards01-2m fp32_vocab_reduction: ESTIMATOR DEFECT, disclosed 2026-08-31 (P1-06). The scorer computed the vocabulary reduction in float32 and cast the finished sum to float64; this row previously declared accumulation_dtype float64. Relabeled float32_reduce_legacy -- the value is unchanged, the comparability key moved, and the row ranks only against rows from the same float32-reducing scorer. Synthetic worst case for the defect class: negative per-token 'KL' near -1e-6 against a true value of ~2e-8 on near-equal distributions; this ladder's published means sit at 1e-3..1e-1, three to five orders above that error scale. See docs/PUBLISHED-CORRECTIONS.md.

Panel: malaiwah Qwen3.8-27B suite v5, shard 0 -- 512 contexts

Derived from panel--qwen38.malaiwah.suite-v5-10m by shard_subset: shard 0 of 10 (512 of 5,120 contexts, 330 of 842 source clusters). Different tokens, therefore a different digest and a different comparability key. K6-parity 0.001634 lives here; the FP8 baseline on this panel is 0.005197, NOT the 10M panel's 0.005294.

Panel disclosure -- unsealed_source: The qwen38 v5 token suite is pinned by suite_token_sha256 and by its manifest digest c79dfad3..., but the token files themselves are not published, so a third party cannot reproduce the digest today.

This panel carries 3 separate comparability groups. They are different measurements of different things and are never merged.

Group cmp--75b64be1f101ed22 -- 12 rows

Panel panel--qwen38.malaiwah.suite-v5-shard0-1m -- malaiwah Qwen3.8-27B suite v5, shard 0 -- 512 contexts 512 contexts x 2047 scored positions = 1,048,064 scored positions, score_from 0 sealed: yes (token digest caef8a4628d6c07c...) -- contamination scan: yes, 0 hits Reference (teacher) reference--malaiwah.qwen38-bf16-vllm.suite-v5-shard0-1m -- native_bf16, artifact artifact--qwen.qwen3.8-27b-bf16 @unpinned revision Metric mean_tokenwise_kld, direction reference_to_candidate, accumulation float32_reduce_legacy Estimation surface stack_relation same_stack, head_policy shared_reference_head Comparability key cmp--75b64be1f101ed22 Like-for-like predicate comparable: false -- a RECORDED secondary dimension differs across members: scope. Equal keys make these rows candidates for comparison, not certified like-for-like; ranking across the differing dimension attributes a lane/pipeline/hardware/scope effect to quantization quality. Machine-readable form with per-dimension values: this key's comparability block in index.json.

What this table is. Every row here shares the comparability key above: the same tokens, the same teacher capture, the same metric and direction, the same estimator precision, the same stack relation and the same head policy. That makes them CANDIDATES for ranking -- the key is a necessary partition, not a certificate. Whether they are also like-for-like on the dimensions the key omits (lane, pipeline, scope coverage, hardware) is what the predicate line above answers.

Rank is not a verdict. The table is sorted by fidelity alone, and fidelity buys bits: a larger, higher-bitrate quant will usually sit above a smaller one, which is not news. Read the Size and Codec columns before reading the order, and compare like against like.

What it is NOT comparable to. The nearest neighbouring groups differ in:

• cmp--35a4b2ab8ed5cd50 (4 rows): stack_relation same_stack -> cross_stack
• cmp--05e16411a5932713 (3 rows): reference_id reference--malaiwah.qwen38-bf16-vllm.suite-v5-shard0-1m -> reference--malaiwah.qwen38-bf16-hf.suite-v5-shard0-1m; accumulation_dtype float32_reduce_legacy -> float64
• cmp--c8c4df32774bdb63 (6 rows): panel_id panel--qwen38.malaiwah.suite-v5-shard0-1m -> panel--qwen38.malaiwah.suite-v5-10m; reference_id reference--malaiwah.qwen38-bf16-vllm.suite-v5-shard0-1m -> reference--malaiwah.qwen38-bf16-vllm.suite-v5-10m
• cmp--726ac1b18b8129fa (5 rows): panel_id panel--qwen38.malaiwah.suite-v5-shard0-1m -> panel--qwen38.malaiwah.suite-v5-shard0-1m.scorefrom1024; reference_id reference--malaiwah.qwen38-bf16-vllm.suite-v5-shard0-1m -> reference--malaiwah.qwen38-bf16-vllm.suite-v5-shard0-1m.scorefrom1024

Those numbers are in this file, under their own headings. Quoting one under the other heading is the mistake this layout exists to prevent: the key is a function of the panel, the teacher, the metric, the direction and the estimator, and the validator recomputes it from those fields rather than trusting the stamped value. What that catches is a row filed under a key its own fields do not produce. It does not catch a number attributed to the wrong panel in the first place -- no offline checker can. That is what the receipt digests on every row are for.

Also, and always: every table for a different model. A KL number is a divergence over one model's own vocabulary against that model's own teacher. It is not a quality score that can be carried between models.

Artifact	Codec	Size	mean_tokenwise_kld (nats)	CI95	Top-1	Runs	Attribution	Receipt
turboderp Qwen3.8-27B exl3 6.00bpw	exl3-mcg @6	23.0 GB	0.00158316	[0.00149496, 0.00168324]	98.28 %	1 run, unevidenced	measured by us (their artifact)	receipt
malaiwah Qwen3.8-27B EXL3 K6-parity	exl3-mcg @6	23.1 GB	0.00163382	[0.00154118, 0.00174151]	98.25 %	1 run, unevidenced	measured by us	receipt
malaiwah Qwen3.8-27B EXL3 K5K6 hydrated	exl3-mcg @5	21.6 GB	0.00269988	[0.00251653, 0.00291183]	97.80 %	1 run, unevidenced	measured by us	receipt
malaiwah Qwen3.8-27B EXL3 K5K6	exl3-mcg @5	30.6 GB	0.00314136	[0.00294675, 0.00336868]	97.61 %	1 run, unevidenced	measured by us	receipt
malaiwah Qwen3.8-27B EXL3 K5K6 context	exl3-mcg @5	20.7 GB	0.00340941	[0.00317041, 0.00368653]	97.55 %	1 run, unevidenced	measured by us	receipt
turboderp Qwen3.8-27B exl3 5.00bpw	exl3-mcg @5	19.9 GB	0.00400463	[0.00371442, 0.00433631]	97.37 %	1 run, unevidenced	measured by us (their artifact)	receipt
Qwen3.8-27B FP8 (official)	fp8_e4m3 @8	30.9 GB	0.00519706	[0.00487991, 0.00555746]	96.92 %	1 run, unevidenced	measured by us (their artifact)	receipt
malaiwah Qwen3.8-27B K4	exl3-mcg @4	28.3 GB	0.0103453	[0.00956259, 0.0112458]	95.91 %	1 run, unevidenced	measured by us	receipt
Qwen3.8-27B AWQ-INT4 (upstream unattributed)	awq @4	--	0.0228179	[0.0212457, 0.024624]	93.94 %	1 run, unevidenced	measured by us	receipt
unsloth Qwen3.8-27B NVFP4	nvfp4 @4	--	0.0301154	[0.0276372, 0.0329647]	93.16 %	1 run, unevidenced	measured by us (their artifact)	receipt
gittensor-model-hub Qwen3.8-27B NVFP4 (RTX5090)	nvfp4 @4	20.6 GB	0.0621631	[0.0584911, 0.0663596]	89.85 %	1 run, unevidenced	measured by us (their artifact)	receipt
Qwen3.8-27B MTP-NVFP4 (upstream unattributed)	nvfp4 @4	--	0.15128	[0.141538, 0.163025]	84.74 %	1 run, unevidenced	measured by us	receipt

The same artifact, measured elsewhere in this file. 6 of the artifacts below also carry a number in another table -- on a different panel, teacher or estimator -- and the widest of those spans 77%. None of the readings is wrong and none is interchangeable with another. Quoting one of them as the number for the artifact, without its table, is the misuse this registry exists to make obvious.

the 6 artifacts and their ranges

• Qwen3.8-27B FP8 (official) -- 6 values here, from 0.00298985 to 0.00529563 nats (77% apart). Other tables: cmp--05e16411a5932713, cmp--0bb49e8411b6dc75, cmp--47c0bc74ebec3fa7, cmp--726ac1b18b8129fa, cmp--c8c4df32774bdb63.
• malaiwah Qwen3.8-27B EXL3 K5K6 -- 5 values here, from 0.0030196 to 0.00320988 nats (6% apart). Other tables: cmp--0bb49e8411b6dc75, cmp--47c0bc74ebec3fa7, cmp--726ac1b18b8129fa, cmp--c8c4df32774bdb63.
• malaiwah Qwen3.8-27B EXL3 K5K6 context -- 5 values here, from 0.00324322 to 0.00350936 nats (8% apart). Other tables: cmp--0bb49e8411b6dc75, cmp--47c0bc74ebec3fa7, cmp--726ac1b18b8129fa, cmp--c8c4df32774bdb63.
• malaiwah Qwen3.8-27B EXL3 K5K6 hydrated -- 5 values here, from 0.00257964 to 0.00275963 nats (7% apart). Other tables: cmp--0bb49e8411b6dc75, cmp--47c0bc74ebec3fa7, cmp--726ac1b18b8129fa, cmp--c8c4df32774bdb63.
• malaiwah Qwen3.8-27B K4 -- 5 values here, from 0.00987561 to 0.0106039 nats (7% apart). Other tables: cmp--0bb49e8411b6dc75, cmp--47c0bc74ebec3fa7, cmp--726ac1b18b8129fa, cmp--c8c4df32774bdb63.
• unsloth Qwen3.8-27B NVFP4 -- 2 values here, from 0.0301154 to 0.0310586 nats (3% apart). Other tables: cmp--c8c4df32774bdb63.

Disclosures for the rows above (42)

• qwen38.turboderp-6bpw.suite-v5-shard0-1m revision_unpinned: No measurement receipt for this artifact records a Hub revision. Every kld5 receipt records model_revision=null / model_revision_source='none'. Identity rests on index_sha256 and the per-shard sha256 map the receipt carries.
• qwen38.turboderp-6bpw.suite-v5-shard0-1m artifact_identity_incomplete: The per-tensor-class quantization recipe for this artifact was never published, so scope.assignments records 'unknown' rather than a guessed allocation. Its scope_digest shows the gap.
• qwen38.turboderp-6bpw.suite-v5-shard0-1m single_run: One pass. Repeatability was not established for this row.
• qwen38.turboderp-6bpw.suite-v5-shard0-1m fp32_vocab_reduction: ESTIMATOR DEFECT, disclosed 2026-08-31 (P1-06). The scorer computed the vocabulary reduction in float32 and cast the finished sum to float64; this row previously declared accumulation_dtype float64. Relabeled float32_reduce_legacy -- the value is unchanged, the comparability key moved, and the row ranks only against rows from the same float32-reducing scorer. Synthetic worst case for the defect class: negative per-token 'KL' near -1e-6 against a true value of ~2e-8 on near-equal distributions; this ladder's published means sit at 1e-3..1e-1, three to five orders above that error scale. See docs/PUBLISHED-CORRECTIONS.md.
• qwen38.k6-parity.suite-v5-shard0-1m revision_unpinned: No measurement receipt for this artifact records a Hub revision. Every kld5 receipt records model_revision=null / model_revision_source='none'. Identity rests on index_sha256 and the per-shard sha256 map the receipt carries.
• qwen38.k6-parity.suite-v5-shard0-1m single_run: One pass. Repeatability was not established for this row.
• qwen38.k6-parity.suite-v5-shard0-1m fp32_vocab_reduction: ESTIMATOR DEFECT, disclosed 2026-08-31 (P1-06). The scorer computed the vocabulary reduction in float32 and cast the finished sum to float64; this row previously declared accumulation_dtype float64. Relabeled float32_reduce_legacy -- the value is unchanged, the comparability key moved, and the row ranks only against rows from the same float32-reducing scorer. Synthetic worst case for the defect class: negative per-token 'KL' near -1e-6 against a true value of ~2e-8 on near-equal distributions; this ladder's published means sit at 1e-3..1e-1, three to five orders above that error scale. See docs/PUBLISHED-CORRECTIONS.md.
• qwen38.k5k6-hydrated.suite-v5-shard0-1m revision_unpinned: No measurement receipt for this artifact records a Hub revision. Every kld5 receipt records model_revision=null / model_revision_source='none'. Identity rests on index_sha256 and the per-shard sha256 map the receipt carries.
• qwen38.k5k6-hydrated.suite-v5-shard0-1m single_run: One pass. Repeatability was not established for this row.
• qwen38.k5k6-hydrated.suite-v5-shard0-1m fp32_vocab_reduction: ESTIMATOR DEFECT, disclosed 2026-08-31 (P1-06). The scorer computed the vocabulary reduction in float32 and cast the finished sum to float64; this row previously declared accumulation_dtype float64. Relabeled float32_reduce_legacy -- the value is unchanged, the comparability key moved, and the row ranks only against rows from the same float32-reducing scorer. Synthetic worst case for the defect class: negative per-token 'KL' near -1e-6 against a true value of ~2e-8 on near-equal distributions; this ladder's published means sit at 1e-3..1e-1, three to five orders above that error scale. See docs/PUBLISHED-CORRECTIONS.md.
• qwen38.k5k6.suite-v5-shard0-1m revision_unpinned: No measurement receipt for this artifact records a Hub revision. Every kld5 receipt records model_revision=null / model_revision_source='none'. Identity rests on index_sha256 and the per-shard sha256 map the receipt carries.
• qwen38.k5k6.suite-v5-shard0-1m single_run: One pass. Repeatability was not established for this row.
• qwen38.k5k6.suite-v5-shard0-1m fp32_vocab_reduction: ESTIMATOR DEFECT, disclosed 2026-08-31 (P1-06). The scorer computed the vocabulary reduction in float32 and cast the finished sum to float64; this row previously declared accumulation_dtype float64. Relabeled float32_reduce_legacy -- the value is unchanged, the comparability key moved, and the row ranks only against rows from the same float32-reducing scorer. Synthetic worst case for the defect class: negative per-token 'KL' near -1e-6 against a true value of ~2e-8 on near-equal distributions; this ladder's published means sit at 1e-3..1e-1, three to five orders above that error scale. See docs/PUBLISHED-CORRECTIONS.md.
• qwen38.k5k6-context.suite-v5-shard0-1m revision_unpinned: No measurement receipt for this artifact records a Hub revision. Every kld5 receipt records model_revision=null / model_revision_source='none'. Identity rests on index_sha256 and the per-shard sha256 map the receipt carries.
• qwen38.k5k6-context.suite-v5-shard0-1m single_run: One pass. Repeatability was not established for this row.
• qwen38.k5k6-context.suite-v5-shard0-1m fp32_vocab_reduction: ESTIMATOR DEFECT, disclosed 2026-08-31 (P1-06). The scorer computed the vocabulary reduction in float32 and cast the finished sum to float64; this row previously declared accumulation_dtype float64. Relabeled float32_reduce_legacy -- the value is unchanged, the comparability key moved, and the row ranks only against rows from the same float32-reducing scorer. Synthetic worst case for the defect class: negative per-token 'KL' near -1e-6 against a true value of ~2e-8 on near-equal distributions; this ladder's published means sit at 1e-3..1e-1, three to five orders above that error scale. See docs/PUBLISHED-CORRECTIONS.md.
• qwen38.turboderp-5bpw.suite-v5-shard0-1m revision_unpinned: No measurement receipt for this artifact records a Hub revision. Every kld5 receipt records model_revision=null / model_revision_source='none'. Identity rests on index_sha256 and the per-shard sha256 map the receipt carries.
• qwen38.turboderp-5bpw.suite-v5-shard0-1m artifact_identity_incomplete: The per-tensor-class quantization recipe for this artifact was never published, so scope.assignments records 'unknown' rather than a guessed allocation. Its scope_digest shows the gap.
• qwen38.turboderp-5bpw.suite-v5-shard0-1m single_run: One pass. Repeatability was not established for this row.
• qwen38.turboderp-5bpw.suite-v5-shard0-1m fp32_vocab_reduction: ESTIMATOR DEFECT, disclosed 2026-08-31 (P1-06). The scorer computed the vocabulary reduction in float32 and cast the finished sum to float64; this row previously declared accumulation_dtype float64. Relabeled float32_reduce_legacy -- the value is unchanged, the comparability key moved, and the row ranks only against rows from the same float32-reducing scorer. Synthetic worst case for the defect class: negative per-token 'KL' near -1e-6 against a true value of ~2e-8 on near-equal distributions; this ladder's published means sit at 1e-3..1e-1, three to five orders above that error scale. See docs/PUBLISHED-CORRECTIONS.md.
• qwen38.official-fp8.suite-v5-shard0-1m revision_unpinned: No measurement receipt for this artifact records a Hub revision. Every kld5 receipt records model_revision=null / model_revision_source='none'. Identity rests on index_sha256 and the per-shard sha256 map the receipt carries.
• qwen38.official-fp8.suite-v5-shard0-1m single_run: One pass. Repeatability was not established for this row.
• qwen38.official-fp8.suite-v5-shard0-1m fp32_vocab_reduction: ESTIMATOR DEFECT, disclosed 2026-08-31 (P1-06). The scorer computed the vocabulary reduction in float32 and cast the finished sum to float64; this row previously declared accumulation_dtype float64. Relabeled float32_reduce_legacy -- the value is unchanged, the comparability key moved, and the row ranks only against rows from the same float32-reducing scorer. Synthetic worst case for the defect class: negative per-token 'KL' near -1e-6 against a true value of ~2e-8 on near-equal distributions; this ladder's published means sit at 1e-3..1e-1, three to five orders above that error scale. See docs/PUBLISHED-CORRECTIONS.md.
• qwen38.k4.suite-v5-shard0-1m revision_unpinned: No measurement receipt for this artifact records a Hub revision. Every kld5 receipt records model_revision=null / model_revision_source='none'. Identity rests on index_sha256 and the per-shard sha256 map the receipt carries.
• qwen38.k4.suite-v5-shard0-1m single_run: One pass. Repeatability was not established for this row.
• qwen38.k4.suite-v5-shard0-1m fp32_vocab_reduction: ESTIMATOR DEFECT, disclosed 2026-08-31 (P1-06). The scorer computed the vocabulary reduction in float32 and cast the finished sum to float64; this row previously declared accumulation_dtype float64. Relabeled float32_reduce_legacy -- the value is unchanged, the comparability key moved, and the row ranks only against rows from the same float32-reducing scorer. Synthetic worst case for the defect class: negative per-token 'KL' near -1e-6 against a true value of ~2e-8 on near-equal distributions; this ladder's published means sit at 1e-3..1e-1, three to five orders above that error scale. See docs/PUBLISHED-CORRECTIONS.md.
• qwen38.awq-int4.suite-v5-shard0-1m revision_unpinned: No measurement receipt for this artifact records a Hub revision. Every kld5 receipt records model_revision=null / model_revision_source='none'. Identity rests on index_sha256 and the per-shard sha256 map the receipt carries.
• qwen38.awq-int4.suite-v5-shard0-1m artifact_identity_incomplete: The upstream repository for this artifact is not recorded by the receipt; only a local path. The measurement is ours and real, the artifact identity is not established.
• qwen38.awq-int4.suite-v5-shard0-1m single_run: One pass. Repeatability was not established for this row.
• qwen38.awq-int4.suite-v5-shard0-1m fp32_vocab_reduction: ESTIMATOR DEFECT, disclosed 2026-08-31 (P1-06). The scorer computed the vocabulary reduction in float32 and cast the finished sum to float64; this row previously declared accumulation_dtype float64. Relabeled float32_reduce_legacy -- the value is unchanged, the comparability key moved, and the row ranks only against rows from the same float32-reducing scorer. Synthetic worst case for the defect class: negative per-token 'KL' near -1e-6 against a true value of ~2e-8 on near-equal distributions; this ladder's published means sit at 1e-3..1e-1, three to five orders above that error scale. See docs/PUBLISHED-CORRECTIONS.md.
• qwen38.unsloth-nvfp4.suite-v5-shard0-1m revision_unpinned: No measurement receipt for this artifact records a Hub revision. Every kld5 receipt records model_revision=null / model_revision_source='none'. Identity rests on index_sha256 and the per-shard sha256 map the receipt carries.
• qwen38.unsloth-nvfp4.suite-v5-shard0-1m artifact_identity_incomplete: The per-tensor-class quantization recipe for this artifact was never published, so scope.assignments records 'unknown' rather than a guessed allocation. Its scope_digest shows the gap.
• qwen38.unsloth-nvfp4.suite-v5-shard0-1m single_run: One pass. Repeatability was not established for this row.
• qwen38.unsloth-nvfp4.suite-v5-shard0-1m fp32_vocab_reduction: ESTIMATOR DEFECT, disclosed 2026-08-31 (P1-06). The scorer computed the vocabulary reduction in float32 and cast the finished sum to float64; this row previously declared accumulation_dtype float64. Relabeled float32_reduce_legacy -- the value is unchanged, the comparability key moved, and the row ranks only against rows from the same float32-reducing scorer. Synthetic worst case for the defect class: negative per-token 'KL' near -1e-6 against a true value of ~2e-8 on near-equal distributions; this ladder's published means sit at 1e-3..1e-1, three to five orders above that error scale. See docs/PUBLISHED-CORRECTIONS.md.
• qwen38.gittensor-nvfp4.suite-v5-shard0-1m revision_unpinned: No measurement receipt for this artifact records a Hub revision. Every kld5 receipt records model_revision=null / model_revision_source='none'. Identity rests on index_sha256 and the per-shard sha256 map the receipt carries.
• qwen38.gittensor-nvfp4.suite-v5-shard0-1m artifact_identity_incomplete: The per-tensor-class quantization recipe for this artifact was never published, so scope.assignments records 'unknown' rather than a guessed allocation. Its scope_digest shows the gap.
• qwen38.gittensor-nvfp4.suite-v5-shard0-1m single_run: One pass. Repeatability was not established for this row.
• qwen38.gittensor-nvfp4.suite-v5-shard0-1m fp32_vocab_reduction: ESTIMATOR DEFECT, disclosed 2026-08-31 (P1-06). The scorer computed the vocabulary reduction in float32 and cast the finished sum to float64; this row previously declared accumulation_dtype float64. Relabeled float32_reduce_legacy -- the value is unchanged, the comparability key moved, and the row ranks only against rows from the same float32-reducing scorer. Synthetic worst case for the defect class: negative per-token 'KL' near -1e-6 against a true value of ~2e-8 on near-equal distributions; this ladder's published means sit at 1e-3..1e-1, three to five orders above that error scale. See docs/PUBLISHED-CORRECTIONS.md.
• qwen38.mtp-nvfp4.suite-v5-shard0-1m revision_unpinned: No measurement receipt for this artifact records a Hub revision. Every kld5 receipt records model_revision=null / model_revision_source='none'. Identity rests on index_sha256 and the per-shard sha256 map the receipt carries.
• qwen38.mtp-nvfp4.suite-v5-shard0-1m artifact_identity_incomplete: The upstream repository for this artifact is not recorded by the receipt; only a local path. The measurement is ours and real, the artifact identity is not established.
• qwen38.mtp-nvfp4.suite-v5-shard0-1m single_run: One pass. Repeatability was not established for this row.
• qwen38.mtp-nvfp4.suite-v5-shard0-1m fp32_vocab_reduction: ESTIMATOR DEFECT, disclosed 2026-08-31 (P1-06). The scorer computed the vocabulary reduction in float32 and cast the finished sum to float64; this row previously declared accumulation_dtype float64. Relabeled float32_reduce_legacy -- the value is unchanged, the comparability key moved, and the row ranks only against rows from the same float32-reducing scorer. Synthetic worst case for the defect class: negative per-token 'KL' near -1e-6 against a true value of ~2e-8 on near-equal distributions; this ladder's published means sit at 1e-3..1e-1, three to five orders above that error scale. See docs/PUBLISHED-CORRECTIONS.md.

Group cmp--35a4b2ab8ed5cd50 -- 4 rows

Panel panel--qwen38.malaiwah.suite-v5-shard0-1m -- malaiwah Qwen3.8-27B suite v5, shard 0 -- 512 contexts 512 contexts x 2047 scored positions = 1,048,064 scored positions, score_from 0 sealed: yes (token digest caef8a4628d6c07c...) -- contamination scan: yes, 0 hits Reference (teacher) reference--malaiwah.qwen38-bf16-vllm.suite-v5-shard0-1m -- native_bf16, artifact artifact--qwen.qwen3.8-27b-bf16 @unpinned revision Metric mean_tokenwise_kld, direction reference_to_candidate, accumulation float32_reduce_legacy Estimation surface stack_relation cross_stack, head_policy shared_reference_head Comparability key cmp--35a4b2ab8ed5cd50 Like-for-like predicate comparable: unknown -- no recorded difference, but harness, lane, replay_backend, replay_env, scope, stack are unrecorded for at least one member, so homogeneity cannot be certified. Machine-readable form with per-dimension values: this key's comparability block in index.json.

What this table is. Every row here shares the comparability key above: the same tokens, the same teacher capture, the same metric and direction, the same estimator precision, the same stack relation and the same head policy. That makes them CANDIDATES for ranking -- the key is a necessary partition, not a certificate. Whether they are also like-for-like on the dimensions the key omits (lane, pipeline, scope coverage, hardware) is what the predicate line above answers.

Rank is not a verdict. The table is sorted by fidelity alone, and fidelity buys bits: a larger, higher-bitrate quant will usually sit above a smaller one, which is not news. Read the Size and Codec columns before reading the order, and compare like against like.

What it is NOT comparable to. The nearest neighbouring groups differ in:

• cmp--75b64be1f101ed22 (12 rows): stack_relation cross_stack -> same_stack

Those numbers are in this file, under their own headings. Quoting one under the other heading is the mistake this layout exists to prevent: the key is a function of the panel, the teacher, the metric, the direction and the estimator, and the validator recomputes it from those fields rather than trusting the stamped value. What that catches is a row filed under a key its own fields do not produce. It does not catch a number attributed to the wrong panel in the first place -- no offline checker can. That is what the receipt digests on every row are for.

Also, and always: every table for a different model. A KL number is a divergence over one model's own vocabulary against that model's own teacher. It is not a quality score that can be carried between models.

Artifact	Codec	Size	mean_tokenwise_kld (nats)	CI95	Top-1	Runs	Attribution	Receipt
unsloth Qwen3.8-27B-GGUF BF16 (measurement floor)	bf16	54.7 GB	0.000507355	[0.000492078, 0.00052326]	99.07 %	1 run, unevidenced	measured by us (their artifact)	receipt
unsloth Qwen3.8-27B-GGUF Q8_0	gguf-k-quant @8	29.0 GB	0.00108681	[0.00105026, 0.00112685]	98.53 %	1 run, unevidenced	measured by us (their artifact)	receipt
unsloth Qwen3.8-27B-GGUF Q6_K	gguf-k-quant @6	22.9 GB	0.00203522	[0.00193876, 0.00214482]	97.98 %	1 run, unevidenced	measured by us (their artifact)	receipt
unsloth Qwen3.8-27B-GGUF UD-Q5_K_XL	gguf-k-quant @5	20.2 GB	0.00444353	[0.00415816, 0.00476989]	97.20 %	1 run, unevidenced	measured by us (their artifact)	receipt

Bias on unsloth Qwen3.8-27B-GGUF BF16 -- cross_stack_capture_replay, direction upward. THIS ROW IS THE FLOOR. Unquantized BF16 weights read by llama.cpp and scored against the vLLM BF16 reference: what two engines disagree by on identical weights. 0.000507 nats, 99.07% top-1. Every GGUF row on this panel contains this term; no EXL3 or FP8 row does.

Bias on unsloth Qwen3.8-27B-GGUF Q8_0 -- cross_stack_capture_replay, direction unknown. llama.cpp candidate capture vs vLLM reference capture. The cross-engine control on this exact panel is 0.000507 nats. Its signed descriptive difference from this candidate is 0.0005794503201991574; neither a causal allocation nor a bound on serving divergence follows.

Bias on unsloth Qwen3.8-27B-GGUF Q6_K -- cross_stack_capture_replay, direction unknown. llama.cpp candidate capture vs vLLM reference capture. The cross-engine control on this exact panel is 0.000507 nats. Its signed descriptive difference from this candidate is 0.0015278671188742878; neither a causal allocation nor a bound on serving divergence follows.

Bias on unsloth Qwen3.8-27B-GGUF UD-Q5_K_XL -- cross_stack_capture_replay, direction unknown. llama.cpp candidate capture vs vLLM reference capture. The cross-engine control on this exact panel is 0.000507 nats. Its signed descriptive difference from this candidate is 0.003936170795822309; neither a causal allocation nor a bound on serving divergence follows.

Disclosures for the rows above (20)

• qwen38.gguf-bf16-engine-floor.suite-v5-shard0-1m cross_engine_capture: The candidate was captured with llama.cpp; the reference and every EXL3/FP8 row on this panel were captured under vLLM. Runtime and weight differences may amplify or cancel; they are not an additive error budget. The unquantized cross-engine control measured 0.000507 nats.
• qwen38.gguf-bf16-engine-floor.suite-v5-shard0-1m single_run: One pass.
• qwen38.gguf-bf16-engine-floor.suite-v5-shard0-1m fp32_vocab_reduction: ESTIMATOR DEFECT, disclosed 2026-08-31 (P1-06). The scorer computed the vocabulary reduction in float32 and cast the finished sum to float64; this row previously declared accumulation_dtype float64. Relabeled float32_reduce_legacy -- the value is unchanged, the comparability key moved, and the row ranks only against rows from the same float32-reducing scorer. Synthetic worst case for the defect class: negative per-token 'KL' near -1e-6 against a true value of ~2e-8 on near-equal distributions; this ladder's published means sit at 1e-3..1e-1, three to five orders above that error scale. See docs/PUBLISHED-CORRECTIONS.md.
• qwen38.gguf-bf16-engine-floor.suite-v5-shard0-1m revision_unpinned: No measurement receipt for this artifact records a Hub revision. Every kld5 receipt records model_revision=null / model_revision_source='none'. Identity rests on index_sha256 and the per-shard sha256 map the receipt carries.
• qwen38.gguf-bf16-engine-floor.suite-v5-shard0-1m note: CONTROL ROW / CROSS-ENGINE FLOOR.
• qwen38.unsloth-gguf-q8-0.suite-v5-shard0-1m cross_engine_capture: The candidate was captured with llama.cpp; the reference and every EXL3/FP8 row on this panel were captured under vLLM. Runtime and weight differences may amplify or cancel; they are not an additive error budget. The unquantized cross-engine control measured 0.000507 nats.
• qwen38.unsloth-gguf-q8-0.suite-v5-shard0-1m single_run: One pass.
• qwen38.unsloth-gguf-q8-0.suite-v5-shard0-1m fp32_vocab_reduction: ESTIMATOR DEFECT, disclosed 2026-08-31 (P1-06). The scorer computed the vocabulary reduction in float32 and cast the finished sum to float64; this row previously declared accumulation_dtype float64. Relabeled float32_reduce_legacy -- the value is unchanged, the comparability key moved, and the row ranks only against rows from the same float32-reducing scorer. Synthetic worst case for the defect class: negative per-token 'KL' near -1e-6 against a true value of ~2e-8 on near-equal distributions; this ladder's published means sit at 1e-3..1e-1, three to five orders above that error scale. See docs/PUBLISHED-CORRECTIONS.md.
• qwen38.unsloth-gguf-q8-0.suite-v5-shard0-1m revision_unpinned: No measurement receipt for this artifact records a Hub revision. Every kld5 receipt records model_revision=null / model_revision_source='none'. Identity rests on index_sha256 and the per-shard sha256 map the receipt carries.
• qwen38.unsloth-gguf-q8-0.suite-v5-shard0-1m artifact_identity_incomplete: The per-tensor-class quantization recipe for this artifact was never published, so scope.assignments records 'unknown' rather than a guessed allocation. Its scope_digest shows the gap.
• qwen38.unsloth-gguf-q6-k.suite-v5-shard0-1m cross_engine_capture: The candidate was captured with llama.cpp; the reference and every EXL3/FP8 row on this panel were captured under vLLM. Runtime and weight differences may amplify or cancel; they are not an additive error budget. The unquantized cross-engine control measured 0.000507 nats.
• qwen38.unsloth-gguf-q6-k.suite-v5-shard0-1m single_run: One pass.
• qwen38.unsloth-gguf-q6-k.suite-v5-shard0-1m fp32_vocab_reduction: ESTIMATOR DEFECT, disclosed 2026-08-31 (P1-06). The scorer computed the vocabulary reduction in float32 and cast the finished sum to float64; this row previously declared accumulation_dtype float64. Relabeled float32_reduce_legacy -- the value is unchanged, the comparability key moved, and the row ranks only against rows from the same float32-reducing scorer. Synthetic worst case for the defect class: negative per-token 'KL' near -1e-6 against a true value of ~2e-8 on near-equal distributions; this ladder's published means sit at 1e-3..1e-1, three to five orders above that error scale. See docs/PUBLISHED-CORRECTIONS.md.
• qwen38.unsloth-gguf-q6-k.suite-v5-shard0-1m revision_unpinned: No measurement receipt for this artifact records a Hub revision. Every kld5 receipt records model_revision=null / model_revision_source='none'. Identity rests on index_sha256 and the per-shard sha256 map the receipt carries.
• qwen38.unsloth-gguf-q6-k.suite-v5-shard0-1m artifact_identity_incomplete: The per-tensor-class quantization recipe for this artifact was never published, so scope.assignments records 'unknown' rather than a guessed allocation. Its scope_digest shows the gap.
• qwen38.unsloth-gguf-ud-q5-k-xl.suite-v5-shard0-1m cross_engine_capture: The candidate was captured with llama.cpp; the reference and every EXL3/FP8 row on this panel were captured under vLLM. Runtime and weight differences may amplify or cancel; they are not an additive error budget. The unquantized cross-engine control measured 0.000507 nats.
• qwen38.unsloth-gguf-ud-q5-k-xl.suite-v5-shard0-1m single_run: One pass.
• qwen38.unsloth-gguf-ud-q5-k-xl.suite-v5-shard0-1m fp32_vocab_reduction: ESTIMATOR DEFECT, disclosed 2026-08-31 (P1-06). The scorer computed the vocabulary reduction in float32 and cast the finished sum to float64; this row previously declared accumulation_dtype float64. Relabeled float32_reduce_legacy -- the value is unchanged, the comparability key moved, and the row ranks only against rows from the same float32-reducing scorer. Synthetic worst case for the defect class: negative per-token 'KL' near -1e-6 against a true value of ~2e-8 on near-equal distributions; this ladder's published means sit at 1e-3..1e-1, three to five orders above that error scale. See docs/PUBLISHED-CORRECTIONS.md.
• qwen38.unsloth-gguf-ud-q5-k-xl.suite-v5-shard0-1m revision_unpinned: No measurement receipt for this artifact records a Hub revision. Every kld5 receipt records model_revision=null / model_revision_source='none'. Identity rests on index_sha256 and the per-shard sha256 map the receipt carries.
• qwen38.unsloth-gguf-ud-q5-k-xl.suite-v5-shard0-1m artifact_identity_incomplete: The per-tensor-class quantization recipe for this artifact was never published, so scope.assignments records 'unknown' rather than a guessed allocation. Its scope_digest shows the gap.

Group cmp--05e16411a5932713 -- 3 rows

Panel panel--qwen38.malaiwah.suite-v5-shard0-1m -- malaiwah Qwen3.8-27B suite v5, shard 0 -- 512 contexts 512 contexts x 2047 scored positions = 1,048,064 scored positions, score_from 0 sealed: yes (token digest caef8a4628d6c07c...) -- contamination scan: yes, 0 hits Reference (teacher) reference--malaiwah.qwen38-bf16-hf.suite-v5-shard0-1m -- native_bf16, artifact artifact--qwen.qwen3.8-27b-bf16 @unpinned revision Metric mean_tokenwise_kld, direction reference_to_candidate, accumulation float64 Estimation surface stack_relation same_stack, head_policy shared_reference_head Comparability key cmp--05e16411a5932713 Like-for-like predicate comparable: unknown -- no recorded difference, but replay_backend, replay_env, stack are unrecorded for at least one member, so homogeneity cannot be certified. Machine-readable form with per-dimension values: this key's comparability block in index.json.

What this table is. Every row here shares the comparability key above: the same tokens, the same teacher capture, the same metric and direction, the same estimator precision, the same stack relation and the same head policy. That makes them CANDIDATES for ranking -- the key is a necessary partition, not a certificate. Whether they are also like-for-like on the dimensions the key omits (lane, pipeline, scope coverage, hardware) is what the predicate line above answers.

Rank is not a verdict. The table is sorted by fidelity alone, and fidelity buys bits: a larger, higher-bitrate quant will usually sit above a smaller one, which is not news. Read the Size and Codec columns before reading the order, and compare like against like.

What it is NOT comparable to. The nearest neighbouring groups differ in:

• cmp--75b64be1f101ed22 (12 rows): reference_id reference--malaiwah.qwen38-bf16-hf.suite-v5-shard0-1m -> reference--malaiwah.qwen38-bf16-vllm.suite-v5-shard0-1m; accumulation_dtype float64 -> float32_reduce_legacy

Those numbers are in this file, under their own headings. Quoting one under the other heading is the mistake this layout exists to prevent: the key is a function of the panel, the teacher, the metric, the direction and the estimator, and the validator recomputes it from those fields rather than trusting the stamped value. What that catches is a row filed under a key its own fields do not produce. It does not catch a number attributed to the wrong panel in the first place -- no offline checker can. That is what the receipt digests on every row are for.

Also, and always: every table for a different model. A KL number is a divergence over one model's own vocabulary against that model's own teacher. It is not a quality score that can be carried between models.

Artifact	Codec	Size	mean_tokenwise_kld (nats)	CI95	Top-1	Runs	Attribution	Receipt
Qwen3.8-27B BF16 (measurement floor)	bf16	--	0	--	100.00 %	3 runs, bitwise identical	measured by us (their artifact)	receipt
Qwen3.8-27B FP8 (official)	fp8_e4m3 @8	30.9 GB	0.00298985	--	97.75 %	1 run, unevidenced	measured by us (their artifact)	receipt
cyankiwi Qwen3.8-27B AWQ-INT4	awq @4	21.0 GB	0.0224494	--	94.02 %	1 run, unevidenced	measured by us (their artifact)	receipt

The same artifact, measured elsewhere in this file. One of the artifacts below also carries a number in another table -- on a different panel, teacher or estimator -- and the widest of those spans 77%. None of the readings is wrong and none is interchangeable with another. Quoting one of them as the number for the artifact, without its table, is the misuse this registry exists to make obvious.

• Qwen3.8-27B FP8 (official) -- 6 values here, from 0.00298985 to 0.00529563 nats (77% apart). Other tables: cmp--0bb49e8411b6dc75, cmp--47c0bc74ebec3fa7, cmp--726ac1b18b8129fa, cmp--75b64be1f101ed22, cmp--c8c4df32774bdb63.

Disclosures for the rows above (5)

• qwen38-hf.fp8-dequantized.suite-v5-shard0-1m lossy_capture_codec: RECONSTRUCTED, NOT EXECUTED. The vendor FP8 path is unavailable on this hardware: the fused deep-gemm kernel aborts with 'Unknown recipe' on Blackwell. The candidate was therefore captured from a bf16 materialisation of the stored fp8 weights (k6/tools/dequant_fp8.py, w = fp8 * weight_scale_inv over 128x128 blocks, accumulated fp32, stored bf16). This is the dequantize-and-run methodology the GGUF/EXL3/MLX ecosystems use for KLD: it measures the error of the STORED weights, not of the vendor kernel. Validated before use: per-tensor rel-L2 against the root is 0.0265 uniformly across gate/up/down/q projections, which is FP8 E4M3's expected error and confirms the scale convention.
• qwen38-hf.fp8-dequantized.suite-v5-shard0-1m estimator_scope_narrower_than_artifact: WEIGHTS-ONLY, NOT A SERVING BOUND. The checkpoint declares activation_scheme: 'dynamic', so the served model also quantizes activations per-token. That operation is absent here and may amplify or cancel weight differences. This is a different estimand with no guaranteed bias direction. It is in particular NOT the same quantity as measurement--qwen38.fp8.suite-v5-shard0-1m (0.005197), which ran the real kernel on the vLLM lane.
• qwen38-hf.fp8-dequantized.suite-v5-shard0-1m record_note: UPSTREAM LOADER DEFECT, ROUTED AROUND. Capturing this artifact through stock transformers silently loads it WRONG. The producer's modules_to_not_convert lists '...layers.N.mlp.gate' -- a MoE router that does not exist in this dense checkpoint -- and transformers.quantizers.quantizers_utils.should_convert_module tests re.match(key, full_name), which is anchored only at the START, so that pattern ALSO matches '...layers.N.mlp.gate_proj'. Verified against the real tensor names: 65 of 65 gate_proj modules excluded from fp8 conversion, 0 of 65 up_proj. Their fp8 weights load into plain bf16 Linears with the block scale never applied, and the 65 gate_proj.weight_scale_inv tensors drop out of the load as 'unexpected' -- the only signal, and nothing refuses on it. The dequantisation used here applies all 407 block scales, and the resulting checkpoint loads with 0 unexpected / 0 missing / 0 mismatched.
• qwen38-hf.fp8-dequantized.suite-v5-shard0-1m single_run: One cold capture of the candidate. Repeatability was not established for the candidate side. The REFERENCE side is the three-run bitwise-identical capture the floor row uses, and the comparison itself is deterministic offline arithmetic over sealed tensors, so the unrepeated term is the candidate forward pass alone.
• qwen38-hf.awq-int4-cyankiwi.suite-v5-shard0-1m single_run: One cold capture of the candidate. Repeatability was not established for the candidate side. The REFERENCE side is the three-run bitwise-identical capture the floor row uses, and the comparison itself is deterministic offline arithmetic over sealed tensors, so the unrepeated term is the candidate forward pass alone.

Panel: malaiwah Qwen3.8-27B suite v5 shard 0, scored from position 256

Derived from panel--qwen38.malaiwah.suite-v5-shard0-1m by scoring_window_change: score_from 0 -> 256 on shard 0.

Panel disclosure -- unsealed_source: The qwen38 v5 token suite is pinned by suite_token_sha256 and by its manifest digest c79dfad3..., but the token files themselves are not published, so a third party cannot reproduce the digest today.

Group cmp--0bb49e8411b6dc75 -- 5 rows

Panel panel--qwen38.malaiwah.suite-v5-shard0-1m.scorefrom256 -- malaiwah Qwen3.8-27B suite v5 shard 0, scored from position 256 512 contexts x 1791 scored positions = 916,992 scored positions, score_from 256, windowed sealed: yes (token digest caef8a4628d6c07c...) -- contamination scan: yes, 0 hits Reference (teacher) reference--malaiwah.qwen38-bf16-vllm.suite-v5-shard0-1m.scorefrom256 -- native_bf16, artifact artifact--qwen.qwen3.8-27b-bf16 @unpinned revision Metric mean_tokenwise_kld, direction reference_to_candidate, accumulation float32_reduce_legacy Estimation surface stack_relation same_stack, head_policy shared_reference_head Comparability key cmp--0bb49e8411b6dc75 Like-for-like predicate comparable: false -- a RECORDED secondary dimension differs across members: scope. Equal keys make these rows candidates for comparison, not certified like-for-like; ranking across the differing dimension attributes a lane/pipeline/hardware/scope effect to quantization quality. Machine-readable form with per-dimension values: this key's comparability block in index.json.

What this table is. Every row here shares the comparability key above: the same tokens, the same teacher capture, the same metric and direction, the same estimator precision, the same stack relation and the same head policy. That makes them CANDIDATES for ranking -- the key is a necessary partition, not a certificate. Whether they are also like-for-like on the dimensions the key omits (lane, pipeline, scope coverage, hardware) is what the predicate line above answers.

Rank is not a verdict. The table is sorted by fidelity alone, and fidelity buys bits: a larger, higher-bitrate quant will usually sit above a smaller one, which is not news. Read the Size and Codec columns before reading the order, and compare like against like.

What it is NOT comparable to. The nearest neighbouring groups differ in:

• cmp--75b64be1f101ed22 (12 rows): panel_id panel--qwen38.malaiwah.suite-v5-shard0-1m.scorefrom256 -> panel--qwen38.malaiwah.suite-v5-shard0-1m; reference_id reference--malaiwah.qwen38-bf16-vllm.suite-v5-shard0-1m.scorefrom256 -> reference--malaiwah.qwen38-bf16-vllm.suite-v5-shard0-1m
• cmp--c8c4df32774bdb63 (6 rows): panel_id panel--qwen38.malaiwah.suite-v5-shard0-1m.scorefrom256 -> panel--qwen38.malaiwah.suite-v5-10m; reference_id reference--malaiwah.qwen38-bf16-vllm.suite-v5-shard0-1m.scorefrom256 -> reference--malaiwah.qwen38-bf16-vllm.suite-v5-10m
• cmp--726ac1b18b8129fa (5 rows): panel_id panel--qwen38.malaiwah.suite-v5-shard0-1m.scorefrom256 -> panel--qwen38.malaiwah.suite-v5-shard0-1m.scorefrom1024; reference_id reference--malaiwah.qwen38-bf16-vllm.suite-v5-shard0-1m.scorefrom256 -> reference--malaiwah.qwen38-bf16-vllm.suite-v5-shard0-1m.scorefrom1024
• cmp--47c0bc74ebec3fa7 (5 rows): panel_id panel--qwen38.malaiwah.suite-v5-shard0-1m.scorefrom256 -> panel--qwen38.malaiwah.suite-v5-shards01-2m; reference_id reference--malaiwah.qwen38-bf16-vllm.suite-v5-shard0-1m.scorefrom256 -> reference--malaiwah.qwen38-bf16-vllm.suite-v5-shards01-2m

Those numbers are in this file, under their own headings. Quoting one under the other heading is the mistake this layout exists to prevent: the key is a function of the panel, the teacher, the metric, the direction and the estimator, and the validator recomputes it from those fields rather than trusting the stamped value. What that catches is a row filed under a key its own fields do not produce. It does not catch a number attributed to the wrong panel in the first place -- no offline checker can. That is what the receipt digests on every row are for.

Also, and always: every table for a different model. A KL number is a divergence over one model's own vocabulary against that model's own teacher. It is not a quality score that can be carried between models.

Artifact	Codec	Size	mean_tokenwise_kld (nats)	CI95	Top-1	Runs	Attribution	Receipt
malaiwah Qwen3.8-27B EXL3 K5K6 hydrated	exl3-mcg @5	21.6 GB	0.00265978	[0.0024736, 0.00287669]	97.84 %	1 run, unevidenced	measured by us	receipt
malaiwah Qwen3.8-27B EXL3 K5K6	exl3-mcg @5	30.6 GB	0.00310033	[0.00290065, 0.00333055]	97.64 %	1 run, unevidenced	measured by us	receipt
malaiwah Qwen3.8-27B EXL3 K5K6 context	exl3-mcg @5	20.7 GB	0.00334231	[0.00310111, 0.00362553]	97.59 %	1 run, unevidenced	measured by us	receipt
Qwen3.8-27B FP8 (official)	fp8_e4m3 @8	30.9 GB	0.00509007	[0.00477169, 0.00544966]	96.97 %	1 run, unevidenced	measured by us (their artifact)	receipt
malaiwah Qwen3.8-27B K4	exl3-mcg @4	28.3 GB	0.0101538	[0.00936883, 0.0110793]	95.98 %	1 run, unevidenced	measured by us	receipt

The same artifact, measured elsewhere in this file. 5 of the artifacts below also carry a number in another table -- on a different panel, teacher or estimator -- and the widest of those spans 77%. None of the readings is wrong and none is interchangeable with another. Quoting one of them as the number for the artifact, without its table, is the misuse this registry exists to make obvious.

the 5 artifacts and their ranges

• Qwen3.8-27B FP8 (official) -- 6 values here, from 0.00298985 to 0.00529563 nats (77% apart). Other tables: cmp--05e16411a5932713, cmp--47c0bc74ebec3fa7, cmp--726ac1b18b8129fa, cmp--75b64be1f101ed22, cmp--c8c4df32774bdb63.
• malaiwah Qwen3.8-27B EXL3 K5K6 -- 5 values here, from 0.0030196 to 0.00320988 nats (6% apart). Other tables: cmp--47c0bc74ebec3fa7, cmp--726ac1b18b8129fa, cmp--75b64be1f101ed22, cmp--c8c4df32774bdb63.
• malaiwah Qwen3.8-27B EXL3 K5K6 context -- 5 values here, from 0.00324322 to 0.00350936 nats (8% apart). Other tables: cmp--47c0bc74ebec3fa7, cmp--726ac1b18b8129fa, cmp--75b64be1f101ed22, cmp--c8c4df32774bdb63.
• malaiwah Qwen3.8-27B EXL3 K5K6 hydrated -- 5 values here, from 0.00257964 to 0.00275963 nats (7% apart). Other tables: cmp--47c0bc74ebec3fa7, cmp--726ac1b18b8129fa, cmp--75b64be1f101ed22, cmp--c8c4df32774bdb63.
• malaiwah Qwen3.8-27B K4 -- 5 values here, from 0.00987561 to 0.0106039 nats (7% apart). Other tables: cmp--47c0bc74ebec3fa7, cmp--726ac1b18b8129fa, cmp--75b64be1f101ed22, cmp--c8c4df32774bdb63.

Disclosures for the rows above (15)

• qwen38.k5k6-hydrated.suite-v5-shard0-1m.scorefrom256 revision_unpinned: No measurement receipt for this artifact records a Hub revision. Every kld5 receipt records model_revision=null / model_revision_source='none'. Identity rests on index_sha256 and the per-shard sha256 map the receipt carries.
• qwen38.k5k6-hydrated.suite-v5-shard0-1m.scorefrom256 single_run: One pass. Repeatability was not established for this row.
• qwen38.k5k6-hydrated.suite-v5-shard0-1m.scorefrom256 fp32_vocab_reduction: ESTIMATOR DEFECT, disclosed 2026-08-31 (P1-06). The scorer computed the vocabulary reduction in float32 and cast the finished sum to float64; this row previously declared accumulation_dtype float64. Relabeled float32_reduce_legacy -- the value is unchanged, the comparability key moved, and the row ranks only against rows from the same float32-reducing scorer. Synthetic worst case for the defect class: negative per-token 'KL' near -1e-6 against a true value of ~2e-8 on near-equal distributions; this ladder's published means sit at 1e-3..1e-1, three to five orders above that error scale. See docs/PUBLISHED-CORRECTIONS.md.
• qwen38.k5k6.suite-v5-shard0-1m.scorefrom256 revision_unpinned: No measurement receipt for this artifact records a Hub revision. Every kld5 receipt records model_revision=null / model_revision_source='none'. Identity rests on index_sha256 and the per-shard sha256 map the receipt carries.
• qwen38.k5k6.suite-v5-shard0-1m.scorefrom256 single_run: One pass. Repeatability was not established for this row.
• qwen38.k5k6.suite-v5-shard0-1m.scorefrom256 fp32_vocab_reduction: ESTIMATOR DEFECT, disclosed 2026-08-31 (P1-06). The scorer computed the vocabulary reduction in float32 and cast the finished sum to float64; this row previously declared accumulation_dtype float64. Relabeled float32_reduce_legacy -- the value is unchanged, the comparability key moved, and the row ranks only against rows from the same float32-reducing scorer. Synthetic worst case for the defect class: negative per-token 'KL' near -1e-6 against a true value of ~2e-8 on near-equal distributions; this ladder's published means sit at 1e-3..1e-1, three to five orders above that error scale. See docs/PUBLISHED-CORRECTIONS.md.
• qwen38.k5k6-context.suite-v5-shard0-1m.scorefrom256 revision_unpinned: No measurement receipt for this artifact records a Hub revision. Every kld5 receipt records model_revision=null / model_revision_source='none'. Identity rests on index_sha256 and the per-shard sha256 map the receipt carries.
• qwen38.k5k6-context.suite-v5-shard0-1m.scorefrom256 single_run: One pass. Repeatability was not established for this row.
• qwen38.k5k6-context.suite-v5-shard0-1m.scorefrom256 fp32_vocab_reduction: ESTIMATOR DEFECT, disclosed 2026-08-31 (P1-06). The scorer computed the vocabulary reduction in float32 and cast the finished sum to float64; this row previously declared accumulation_dtype float64. Relabeled float32_reduce_legacy -- the value is unchanged, the comparability key moved, and the row ranks only against rows from the same float32-reducing scorer. Synthetic worst case for the defect class: negative per-token 'KL' near -1e-6 against a true value of ~2e-8 on near-equal distributions; this ladder's published means sit at 1e-3..1e-1, three to five orders above that error scale. See docs/PUBLISHED-CORRECTIONS.md.
• qwen38.official-fp8.suite-v5-shard0-1m.scorefrom256 revision_unpinned: No measurement receipt for this artifact records a Hub revision. Every kld5 receipt records model_revision=null / model_revision_source='none'. Identity rests on index_sha256 and the per-shard sha256 map the receipt carries.
• qwen38.official-fp8.suite-v5-shard0-1m.scorefrom256 single_run: One pass. Repeatability was not established for this row.
• qwen38.official-fp8.suite-v5-shard0-1m.scorefrom256 fp32_vocab_reduction: ESTIMATOR DEFECT, disclosed 2026-08-31 (P1-06). The scorer computed the vocabulary reduction in float32 and cast the finished sum to float64; this row previously declared accumulation_dtype float64. Relabeled float32_reduce_legacy -- the value is unchanged, the comparability key moved, and the row ranks only against rows from the same float32-reducing scorer. Synthetic worst case for the defect class: negative per-token 'KL' near -1e-6 against a true value of ~2e-8 on near-equal distributions; this ladder's published means sit at 1e-3..1e-1, three to five orders above that error scale. See docs/PUBLISHED-CORRECTIONS.md.
• qwen38.k4.suite-v5-shard0-1m.scorefrom256 revision_unpinned: No measurement receipt for this artifact records a Hub revision. Every kld5 receipt records model_revision=null / model_revision_source='none'. Identity rests on index_sha256 and the per-shard sha256 map the receipt carries.
• qwen38.k4.suite-v5-shard0-1m.scorefrom256 single_run: One pass. Repeatability was not established for this row.
• qwen38.k4.suite-v5-shard0-1m.scorefrom256 fp32_vocab_reduction: ESTIMATOR DEFECT, disclosed 2026-08-31 (P1-06). The scorer computed the vocabulary reduction in float32 and cast the finished sum to float64; this row previously declared accumulation_dtype float64. Relabeled float32_reduce_legacy -- the value is unchanged, the comparability key moved, and the row ranks only against rows from the same float32-reducing scorer. Synthetic worst case for the defect class: negative per-token 'KL' near -1e-6 against a true value of ~2e-8 on near-equal distributions; this ladder's published means sit at 1e-3..1e-1, three to five orders above that error scale. See docs/PUBLISHED-CORRECTIONS.md.

Panel: malaiwah Qwen3.8-27B suite v5 shard 0, scored from position 1024

Derived from panel--qwen38.malaiwah.suite-v5-shard0-1m by scoring_window_change: score_from 0 -> 1024 on shard 0.

Panel disclosure -- unsealed_source: The qwen38 v5 token suite is pinned by suite_token_sha256 and by its manifest digest c79dfad3..., but the token files themselves are not published, so a third party cannot reproduce the digest today.

Group cmp--726ac1b18b8129fa -- 5 rows

Panel panel--qwen38.malaiwah.suite-v5-shard0-1m.scorefrom1024 -- malaiwah Qwen3.8-27B suite v5 shard 0, scored from position 1024 512 contexts x 1023 scored positions = 523,776 scored positions, score_from 1024, windowed sealed: yes (token digest caef8a4628d6c07c...) -- contamination scan: yes, 0 hits Reference (teacher) reference--malaiwah.qwen38-bf16-vllm.suite-v5-shard0-1m.scorefrom1024 -- native_bf16, artifact artifact--qwen.qwen3.8-27b-bf16 @unpinned revision Metric mean_tokenwise_kld, direction reference_to_candidate, accumulation float32_reduce_legacy Estimation surface stack_relation same_stack, head_policy shared_reference_head Comparability key cmp--726ac1b18b8129fa Like-for-like predicate comparable: false -- a RECORDED secondary dimension differs across members: scope. Equal keys make these rows candidates for comparison, not certified like-for-like; ranking across the differing dimension attributes a lane/pipeline/hardware/scope effect to quantization quality. Machine-readable form with per-dimension values: this key's comparability block in index.json.

What this table is. Every row here shares the comparability key above: the same tokens, the same teacher capture, the same metric and direction, the same estimator precision, the same stack relation and the same head policy. That makes them CANDIDATES for ranking -- the key is a necessary partition, not a certificate. Whether they are also like-for-like on the dimensions the key omits (lane, pipeline, scope coverage, hardware) is what the predicate line above answers.

Rank is not a verdict. The table is sorted by fidelity alone, and fidelity buys bits: a larger, higher-bitrate quant will usually sit above a smaller one, which is not news. Read the Size and Codec columns before reading the order, and compare like against like.

What it is NOT comparable to. The nearest neighbouring groups differ in:

• cmp--75b64be1f101ed22 (12 rows): panel_id panel--qwen38.malaiwah.suite-v5-shard0-1m.scorefrom1024 -> panel--qwen38.malaiwah.suite-v5-shard0-1m; reference_id reference--malaiwah.qwen38-bf16-vllm.suite-v5-shard0-1m.scorefrom1024 -> reference--malaiwah.qwen38-bf16-vllm.suite-v5-shard0-1m
• cmp--c8c4df32774bdb63 (6 rows): panel_id panel--qwen38.malaiwah.suite-v5-shard0-1m.scorefrom1024 -> panel--qwen38.malaiwah.suite-v5-10m; reference_id reference--malaiwah.qwen38-bf16-vllm.suite-v5-shard0-1m.scorefrom1024 -> reference--malaiwah.qwen38-bf16-vllm.suite-v5-10m
• cmp--0bb49e8411b6dc75 (5 rows): panel_id panel--qwen38.malaiwah.suite-v5-shard0-1m.scorefrom1024 -> panel--qwen38.malaiwah.suite-v5-shard0-1m.scorefrom256; reference_id reference--malaiwah.qwen38-bf16-vllm.suite-v5-shard0-1m.scorefrom1024 -> reference--malaiwah.qwen38-bf16-vllm.suite-v5-shard0-1m.scorefrom256
• cmp--47c0bc74ebec3fa7 (5 rows): panel_id panel--qwen38.malaiwah.suite-v5-shard0-1m.scorefrom1024 -> panel--qwen38.malaiwah.suite-v5-shards01-2m; reference_id reference--malaiwah.qwen38-bf16-vllm.suite-v5-shard0-1m.scorefrom1024 -> reference--malaiwah.qwen38-bf16-vllm.suite-v5-shards01-2m

Those numbers are in this file, under their own headings. Quoting one under the other heading is the mistake this layout exists to prevent: the key is a function of the panel, the teacher, the metric, the direction and the estimator, and the validator recomputes it from those fields rather than trusting the stamped value. What that catches is a row filed under a key its own fields do not produce. It does not catch a number attributed to the wrong panel in the first place -- no offline checker can. That is what the receipt digests on every row are for.

Also, and always: every table for a different model. A KL number is a divergence over one model's own vocabulary against that model's own teacher. It is not a quality score that can be carried between models.

Artifact	Codec	Size	mean_tokenwise_kld (nats)	CI95	Top-1	Runs	Attribution	Receipt
malaiwah Qwen3.8-27B EXL3 K5K6 hydrated	exl3-mcg @5	21.6 GB	0.00257964	[0.00239759, 0.00278828]	97.86 %	1 run, unevidenced	measured by us	receipt
malaiwah Qwen3.8-27B EXL3 K5K6	exl3-mcg @5	30.6 GB	0.0030196	[0.0028234, 0.00324563]	97.68 %	1 run, unevidenced	measured by us	receipt
malaiwah Qwen3.8-27B EXL3 K5K6 context	exl3-mcg @5	20.7 GB	0.00324322	[0.00300571, 0.00352013]	97.62 %	1 run, unevidenced	measured by us	receipt
Qwen3.8-27B FP8 (official)	fp8_e4m3 @8	30.9 GB	0.00495487	[0.00463566, 0.005316]	97.02 %	1 run, unevidenced	measured by us (their artifact)	receipt
malaiwah Qwen3.8-27B K4	exl3-mcg @4	28.3 GB	0.00987561	[0.00910329, 0.0107555]	96.04 %	1 run, unevidenced	measured by us	receipt

The same artifact, measured elsewhere in this file. 5 of the artifacts below also carry a number in another table -- on a different panel, teacher or estimator -- and the widest of those spans 77%. None of the readings is wrong and none is interchangeable with another. Quoting one of them as the number for the artifact, without its table, is the misuse this registry exists to make obvious.

the 5 artifacts and their ranges

• Qwen3.8-27B FP8 (official) -- 6 values here, from 0.00298985 to 0.00529563 nats (77% apart). Other tables: cmp--05e16411a5932713, cmp--0bb49e8411b6dc75, cmp--47c0bc74ebec3fa7, cmp--75b64be1f101ed22, cmp--c8c4df32774bdb63.
• malaiwah Qwen3.8-27B EXL3 K5K6 -- 5 values here, from 0.0030196 to 0.00320988 nats (6% apart). Other tables: cmp--0bb49e8411b6dc75, cmp--47c0bc74ebec3fa7, cmp--75b64be1f101ed22, cmp--c8c4df32774bdb63.
• malaiwah Qwen3.8-27B EXL3 K5K6 context -- 5 values here, from 0.00324322 to 0.00350936 nats (8% apart). Other tables: cmp--0bb49e8411b6dc75, cmp--47c0bc74ebec3fa7, cmp--75b64be1f101ed22, cmp--c8c4df32774bdb63.
• malaiwah Qwen3.8-27B EXL3 K5K6 hydrated -- 5 values here, from 0.00257964 to 0.00275963 nats (7% apart). Other tables: cmp--0bb49e8411b6dc75, cmp--47c0bc74ebec3fa7, cmp--75b64be1f101ed22, cmp--c8c4df32774bdb63.
• malaiwah Qwen3.8-27B K4 -- 5 values here, from 0.00987561 to 0.0106039 nats (7% apart). Other tables: cmp--0bb49e8411b6dc75, cmp--47c0bc74ebec3fa7, cmp--75b64be1f101ed22, cmp--c8c4df32774bdb63.

Disclosures for the rows above (15)

• qwen38.k5k6-hydrated.suite-v5-shard0-1m.scorefrom1024 revision_unpinned: No measurement receipt for this artifact records a Hub revision. Every kld5 receipt records model_revision=null / model_revision_source='none'. Identity rests on index_sha256 and the per-shard sha256 map the receipt carries.
• qwen38.k5k6-hydrated.suite-v5-shard0-1m.scorefrom1024 single_run: One pass. Repeatability was not established for this row.
• qwen38.k5k6-hydrated.suite-v5-shard0-1m.scorefrom1024 fp32_vocab_reduction: ESTIMATOR DEFECT, disclosed 2026-08-31 (P1-06). The scorer computed the vocabulary reduction in float32 and cast the finished sum to float64; this row previously declared accumulation_dtype float64. Relabeled float32_reduce_legacy -- the value is unchanged, the comparability key moved, and the row ranks only against rows from the same float32-reducing scorer. Synthetic worst case for the defect class: negative per-token 'KL' near -1e-6 against a true value of ~2e-8 on near-equal distributions; this ladder's published means sit at 1e-3..1e-1, three to five orders above that error scale. See docs/PUBLISHED-CORRECTIONS.md.
• qwen38.k5k6.suite-v5-shard0-1m.scorefrom1024 revision_unpinned: No measurement receipt for this artifact records a Hub revision. Every kld5 receipt records model_revision=null / model_revision_source='none'. Identity rests on index_sha256 and the per-shard sha256 map the receipt carries.
• qwen38.k5k6.suite-v5-shard0-1m.scorefrom1024 single_run: One pass. Repeatability was not established for this row.
• qwen38.k5k6.suite-v5-shard0-1m.scorefrom1024 fp32_vocab_reduction: ESTIMATOR DEFECT, disclosed 2026-08-31 (P1-06). The scorer computed the vocabulary reduction in float32 and cast the finished sum to float64; this row previously declared accumulation_dtype float64. Relabeled float32_reduce_legacy -- the value is unchanged, the comparability key moved, and the row ranks only against rows from the same float32-reducing scorer. Synthetic worst case for the defect class: negative per-token 'KL' near -1e-6 against a true value of ~2e-8 on near-equal distributions; this ladder's published means sit at 1e-3..1e-1, three to five orders above that error scale. See docs/PUBLISHED-CORRECTIONS.md.
• qwen38.k5k6-context.suite-v5-shard0-1m.scorefrom1024 revision_unpinned: No measurement receipt for this artifact records a Hub revision. Every kld5 receipt records model_revision=null / model_revision_source='none'. Identity rests on index_sha256 and the per-shard sha256 map the receipt carries.
• qwen38.k5k6-context.suite-v5-shard0-1m.scorefrom1024 single_run: One pass. Repeatability was not established for this row.
• qwen38.k5k6-context.suite-v5-shard0-1m.scorefrom1024 fp32_vocab_reduction: ESTIMATOR DEFECT, disclosed 2026-08-31 (P1-06). The scorer computed the vocabulary reduction in float32 and cast the finished sum to float64; this row previously declared accumulation_dtype float64. Relabeled float32_reduce_legacy -- the value is unchanged, the comparability key moved, and the row ranks only against rows from the same float32-reducing scorer. Synthetic worst case for the defect class: negative per-token 'KL' near -1e-6 against a true value of ~2e-8 on near-equal distributions; this ladder's published means sit at 1e-3..1e-1, three to five orders above that error scale. See docs/PUBLISHED-CORRECTIONS.md.
• qwen38.official-fp8.suite-v5-shard0-1m.scorefrom1024 revision_unpinned: No measurement receipt for this artifact records a Hub revision. Every kld5 receipt records model_revision=null / model_revision_source='none'. Identity rests on index_sha256 and the per-shard sha256 map the receipt carries.
• qwen38.official-fp8.suite-v5-shard0-1m.scorefrom1024 single_run: One pass. Repeatability was not established for this row.
• qwen38.official-fp8.suite-v5-shard0-1m.scorefrom1024 fp32_vocab_reduction: ESTIMATOR DEFECT, disclosed 2026-08-31 (P1-06). The scorer computed the vocabulary reduction in float32 and cast the finished sum to float64; this row previously declared accumulation_dtype float64. Relabeled float32_reduce_legacy -- the value is unchanged, the comparability key moved, and the row ranks only against rows from the same float32-reducing scorer. Synthetic worst case for the defect class: negative per-token 'KL' near -1e-6 against a true value of ~2e-8 on near-equal distributions; this ladder's published means sit at 1e-3..1e-1, three to five orders above that error scale. See docs/PUBLISHED-CORRECTIONS.md.
• qwen38.k4.suite-v5-shard0-1m.scorefrom1024 revision_unpinned: No measurement receipt for this artifact records a Hub revision. Every kld5 receipt records model_revision=null / model_revision_source='none'. Identity rests on index_sha256 and the per-shard sha256 map the receipt carries.
• qwen38.k4.suite-v5-shard0-1m.scorefrom1024 single_run: One pass. Repeatability was not established for this row.
• qwen38.k4.suite-v5-shard0-1m.scorefrom1024 fp32_vocab_reduction: ESTIMATOR DEFECT, disclosed 2026-08-31 (P1-06). The scorer computed the vocabulary reduction in float32 and cast the finished sum to float64; this row previously declared accumulation_dtype float64. Relabeled float32_reduce_legacy -- the value is unchanged, the comparability key moved, and the row ranks only against rows from the same float32-reducing scorer. Synthetic worst case for the defect class: negative per-token 'KL' near -1e-6 against a true value of ~2e-8 on near-equal distributions; this ladder's published means sit at 1e-3..1e-1, three to five orders above that error scale. See docs/PUBLISHED-CORRECTIONS.md.

GLM-5.2

model--zai-org.glm-5.2 -- published by Z.ai. Tokenizer glm-5.3, vocabulary 154880.

Panel: GLM-5.3 corpus 5-stratum x 5-window panel -- 25 windows x 2048

Panel disclosure -- contamination_unchecked: No overlap scan against GLM-5.3's pretraining data is possible; the five strata are public web text. This affects what the KLD means about the model, not the comparison between two artifacts of it.

Panel disclosure -- small_panel: 25 windows / 51,175 scored positions. On the two 4-bit-class artifacts measured so far the per-window means spread over an order of magnitude (K4: median 0.0030, p95 0.20). Rank artifacts on this panel by the paired per-window difference, never by a single window.

Group cmp--6ccc41df40f849da -- 6 rows

Panel panel--glm53.malaiwah.corpus5x5-v1 -- GLM-5.3 corpus 5-stratum x 5-window panel -- 25 windows x 2048 25 contexts x 2047 scored positions = 51,175 scored positions, score_from 0 sealed: yes (token digest f09ee395f635225a...) -- contamination scan: NOT RUN Reference (teacher) reference--malaiwah.glm-5.2-bf16-hf.corpus5x5-v1 -- native_bf16, artifact artifact--zai-org.glm-5.2-bf16 @cf457fa734ab149ffef225f80893eb38c6ff5cdc Metric mean_tokenwise_kld, direction reference_to_candidate, accumulation float64 Estimation surface stack_relation same_stack, head_policy native_head Comparability key cmp--6ccc41df40f849da Like-for-like predicate comparable: false -- a RECORDED secondary dimension differs across members: scope. Equal keys make these rows candidates for comparison, not certified like-for-like; ranking across the differing dimension attributes a lane/pipeline/hardware/scope effect to quantization quality. Machine-readable form with per-dimension values: this key's comparability block in index.json.

What this table is. Every row here shares the comparability key above: the same tokens, the same teacher capture, the same metric and direction, the same estimator precision, the same stack relation and the same head policy. That makes them CANDIDATES for ranking -- the key is a necessary partition, not a certificate. Whether they are also like-for-like on the dimensions the key omits (lane, pipeline, scope coverage, hardware) is what the predicate line above answers.

Rank is not a verdict. The table is sorted by fidelity alone, and fidelity buys bits: a larger, higher-bitrate quant will usually sit above a smaller one, which is not news. Read the Size and Codec columns before reading the order, and compare like against like.

What it is NOT comparable to. Every other table in this file: no other group shares this key. That includes every table for a different model -- a KL number is a divergence over one model's own vocabulary against that model's own teacher, never a score that can be carried between models.

Artifact	Codec	Size	mean_tokenwise_kld (nats)	CI95	Top-1	Runs	Attribution	Receipt
GLM-5.2 BF16 (the official full-precision release) (measurement floor)	bf16	1506.7 GB	0	--	100.00 %	2 runs, bitwise identical	measured by us (their artifact)	receipt
GLM-5.2 FP8 (the official block-scaled release)	fp8_e4m3 @8	755.6 GB	0.025369	[0.021057, 0.0324859]	95.42 %	2 runs, bitwise identical	measured by us (their artifact)	receipt
unsloth GLM-5.2-GGUF UD-Q4_K_XL (llama.cpp k-quant build, mixed per tensor)	gguf-k-quant @4	467.3 GB	0.0314661	[0.0265774, 0.0394445]	94.99 %	2 runs, bitwise identical	measured by us (their artifact)	receipt
NVIDIA GLM-5.2-NVFP4 (routed experts NVFP4 e2m1 group 16, rest native)	nvfp4 @4	464.8 GB	0.0548369	[0.0450756, 0.069746]	93.44 %	2 runs, bitwise identical	measured by us (their artifact)	receipt
willfalco GLM-5.2-EXL3-TR3-3.25bpw (routed experts trellis mixed K, TP4 rank-sharded, rest native)	exl3-mcg @3.25	339.1 GB	0.0715731	[0.0581496, 0.0899186]	92.69 %	2 runs, bitwise identical	measured by us (their artifact)	receipt
brandonmusic GLM-5.2-EXL3-TR3-3.0bpw (routed experts trellis K3, TP4 rank-sharded, rest native)	exl3-mcg @3	316.4 GB	0.0909455	[0.0730456, 0.120645]	91.74 %	2 runs, bitwise identical	measured by us (their artifact)	receipt

Disclosures for the rows above (13)

• glm-5.2.fp8-dequantized.corpus5x5-v1 lossy_capture_codec: RECONSTRUCTED, NOT EXECUTED. The candidate was captured from a bf16 materialisation of the stored fp8 weights: every fp8_e4m3 tensor is decoded on the host with its 128x128 weight_scale_inv block scale (fp8-block-dequant-to-bf16, accumulated fp32, stored bf16) BEFORE it reaches the loader, so no scale can be silently dropped. This is the dequantize-and-run methodology: it measures the error of the STORED weights, not of a vendor kernel.
• glm-5.2.fp8-dequantized.corpus5x5-v1 activation_quantization_not_captured: WEIGHT-ONLY: the checkpoint declares activation_scheme dynamic, so a served W8A8 deployment also quantizes activations per token at runtime. That term is absent here, so the value is expected to understate the served divergence; it is not a mathematical bound on a mean KL.
• glm-5.2.fp8-dequantized.corpus5x5-v1 note: Per-window mean 0.025368987988553689, population sd 0.013813714764382682, min 0.004672180887474375 (final-0012, literary), max 0.072812005506925861 (final-0014, literary) over 25 windows; the token mean is the published value. NEW GROUP: scored against the GLM-5.2 same-lane root reference--malaiwah.glm-5.2-bf16-hf.corpus5x5-v1, not against the GLM-5.3 root -- do not read it beside a GLM-5.3 row on this panel.
• glm-5.2.gguf-unsloth-udq4kxl.corpus5x5-v1 lossy_capture_codec: RECONSTRUCTED, NOT EXECUTED. Every GGUF tensor is dequantized to bf16 on the capture host (gguf-dequant-to-bf16), k-quant block traits read from the tensor tables themselves. The decoder is proven BITWISE against gguf-py 0.19.0's own gguf.quants.dequantize on real fetched blocks (engines/tools/gguf-evidence/), so the DECODE is not in question. What is absent is the serving engine: llama.cpp runs these weights through its own kernels and its own KV-cache quantization. This row is advisory because it measures the STORED WEIGHTS, not a llama.cpp deployment. There is no activation-quantization caveat: a GGUF k-quant build declares none.
• glm-5.2.gguf-unsloth-udq4kxl.corpus5x5-v1 quantized_head: HEAD-1d with a QUANTIZED head: this build's lm_head is Q8_0, so the candidate replayed through its own dequantized head (b957bc5d2be8...) and the reference through the official bf16 head (a012be05e771...). The head's own quantization error is inside this value -- unlike every other GLM-5.2 row, whose head is the official tensor byte for byte. Read the difference against them as codec-plus-head, not codec alone.
• glm-5.2.gguf-unsloth-udq4kxl.corpus5x5-v1 note: Per-window mean 0.031466114090875824, population sd 0.015674312649714767, min 0.0071216832934662792 (final-0012, literary), max 0.081442432684445523 (final-0014, literary) over 25 windows; the token mean is the published value. NEW GROUP: scored against the GLM-5.2 same-lane root reference--malaiwah.glm-5.2-bf16-hf.corpus5x5-v1, not against the GLM-5.3 root -- do not read it beside a GLM-5.3 row on this panel.
• glm-5.2.nvfp4-nvidia.corpus5x5-v1 lossy_capture_codec: RECONSTRUCTED, NOT EXECUTED. The 57,600 routed-expert NVFP4 component sets are decoded to bf16 per module before the loader (nvfp4-modelopt-dequant-to-bf16: e2m1 unpack, per-group weight_scale in fp8_e4m3 and the per-tensor weight_scale_2, evaluated in exact fp32 and cast once to bf16). The decoder is proven BITWISE against the ecosystem reference implementation -- compressed-tensors 0.18.0's own unpack_fp4_from_uint8 plus the same LUT/scale math in exact fp32 -- on real range-read tensors of THIS release (max_abs_diff_fp32 exactly 0.0, bitwise after the bf16 cast too; engines/tools/nvfp4-evidence/glm53-nvfp4-parity.json). So the decode is not the reason this row is advisory: the reason is that a weights-only capture runs the STORED weights and not a served NVFP4 kernel.
• glm-5.2.nvfp4-nvidia.corpus5x5-v1 activation_quantization_not_captured: WEIGHT-ONLY. The release ships a static per-tensor F32 input_scale beside each of its 57,600 routed-expert modules; a served W4A4 NVFP4 kernel quantizes activations with it and this capture does not. The scales were read and recorded, never applied, so this value is expected to understate a served NVFP4 deployment. It is not a mathematical bound on a mean KL.
• glm-5.2.nvfp4-nvidia.corpus5x5-v1 note: Per-window mean 0.05483693836564809, population sd 0.030292313686806162, min 0.0072093067103579933 (final-0012, literary), max 0.14810523339387668 (final-0014, literary) over 25 windows; the token mean is the published value. NEW GROUP: scored against the GLM-5.2 same-lane root reference--malaiwah.glm-5.2-bf16-hf.corpus5x5-v1, not against the GLM-5.3 root -- do not read it beside a GLM-5.3 row on this panel.
• glm-5.2.exl3-tr3-3.25bpw-willfalco.corpus5x5-v1 lossy_capture_codec: RECONSTRUCTED, NOT EXECUTED. The routed-expert trellis payload groups (57,600 modules x four TP rank shards) are decoded to bf16 per module before the loader by engines/tools/exl3hf_surface.py:decode_payload_hf, this repository's transcription of exllamav3's codebooks, and composed in ascending rank order. That decoder has NOT been proven bitwise against a running exllamav3 kernel -- only against in-house fp64 routes and real payloads -- and the served exllamav3 numerics are not in this number either. Both are why this row is advisory.
• glm-5.2.exl3-tr3-3.25bpw-willfalco.corpus5x5-v1 note: Per-window mean 0.071573100109609725, population sd 0.039097177521238563, min 0.0050477596339557904 (final-0012, literary), max 0.17097192172388301 (final-0014, literary) over 25 windows; the token mean is the published value. NEW GROUP: scored against the GLM-5.2 same-lane root reference--malaiwah.glm-5.2-bf16-hf.corpus5x5-v1, not against the GLM-5.3 root -- do not read it beside a GLM-5.3 row on this panel.
• glm-5.2.exl3-tr3-3.0bpw-brandonmusic.corpus5x5-v1 lossy_capture_codec: RECONSTRUCTED, NOT EXECUTED. The 230,400 trellis payload groups (57,600 modules x four TP rank shards, all K=3) are decoded to bf16 per module before the loader by engines/tools/exl3hf_surface.py:decode_payload_hf, this repository's transcription of exllamav3's mcg codebook, and composed in ascending rank order. That decoder has NOT been proven bitwise against a running exllamav3 kernel -- it is proven against in-house fp64 routes and real payloads only -- and the served exllamav3 numerics are not in this number either. Both are why this row is advisory.
• glm-5.2.exl3-tr3-3.0bpw-brandonmusic.corpus5x5-v1 note: Per-window mean 0.090945547067338733, population sd 0.057483301383525552, min 0.0068051796584907296 (final-0012, literary), max 0.26696687319852314 (final-0014, literary) over 25 windows; the token mean is the published value. NEW GROUP: scored against the GLM-5.2 same-lane root reference--malaiwah.glm-5.2-bf16-hf.corpus5x5-v1, not against the GLM-5.3 root -- do not read it beside a GLM-5.3 row on this panel.

GLM-5.3

model--zai-org.glm-5.3 -- published by Z.ai. Tokenizer glm-5.3, vocabulary 154880.

Panel: GLM-5.3 corpus 5-stratum x 5-window panel -- 25 windows x 2048

Panel disclosure -- contamination_unchecked: No overlap scan against GLM-5.3's pretraining data is possible; the five strata are public web text. This affects what the KLD means about the model, not the comparison between two artifacts of it.

Panel disclosure -- small_panel: 25 windows / 51,175 scored positions. On the two 4-bit-class artifacts measured so far the per-window means spread over an order of magnitude (K4: median 0.0030, p95 0.20). Rank artifacts on this panel by the paired per-window difference, never by a single window.

Group cmp--fdbd312a2551db89 -- 11 rows

Panel panel--glm53.malaiwah.corpus5x5-v1 -- GLM-5.3 corpus 5-stratum x 5-window panel -- 25 windows x 2048 25 contexts x 2047 scored positions = 51,175 scored positions, score_from 0 sealed: yes (token digest f09ee395f635225a...) -- contamination scan: NOT RUN Reference (teacher) reference--malaiwah.glm-5.3-bf16-hf.corpus5x5-v1 -- native_bf16, artifact artifact--zai-org.glm-5.3-bf16 @304b8051cfb2b260b61ce0cbe330e02a98e73639 Metric mean_tokenwise_kld, direction reference_to_candidate, accumulation float64 Estimation surface stack_relation same_stack, head_policy native_head Comparability key cmp--fdbd312a2551db89 Like-for-like predicate comparable: false -- a RECORDED secondary dimension differs across members: scope. Equal keys make these rows candidates for comparison, not certified like-for-like; ranking across the differing dimension attributes a lane/pipeline/hardware/scope effect to quantization quality. Machine-readable form with per-dimension values: this key's comparability block in index.json.

What this table is. Every row here shares the comparability key above: the same tokens, the same teacher capture, the same metric and direction, the same estimator precision, the same stack relation and the same head policy. That makes them CANDIDATES for ranking -- the key is a necessary partition, not a certificate. Whether they are also like-for-like on the dimensions the key omits (lane, pipeline, scope coverage, hardware) is what the predicate line above answers.

Rank is not a verdict. The table is sorted by fidelity alone, and fidelity buys bits: a larger, higher-bitrate quant will usually sit above a smaller one, which is not news. Read the Size and Codec columns before reading the order, and compare like against like.

What it is NOT comparable to. Every other table in this file: no other group shares this key. That includes every table for a different model -- a KL number is a divergence over one model's own vocabulary against that model's own teacher, never a score that can be carried between models.

Artifact	Codec	Size	mean_tokenwise_kld (nats)	CI95	Top-1	Runs	Attribution	Receipt
GLM-5.3 BF16 (the official full-precision release) (measurement floor)	bf16	1506.7 GB	0	--	100.00 %	2 runs, bitwise identical	measured by us (their artifact)	receipt
GLM-5.3 FP8 (the official block-scaled release)	fp8_e4m3 @8	755.6 GB	0.0223051	[0.0189018, 0.0289212]	95.64 %	2 runs, bitwise identical	measured by us (their artifact)	receipt
unsloth GLM-5.3-GGUF UD-Q4_K_XL (llama.cpp k-quant build, mixed per tensor)	gguf-k-quant @4	467.3 GB	0.0275461	[0.0228582, 0.0352615]	95.37 %	2 runs, bitwise identical	measured by us (their artifact)	receipt
wrldsuksgo2mars GLM-5.3 EXL3 K4 v1 (routed experts trellis K4, rest FP8)	exl3-mcg @4	394.0 GB	0.0448038	[0.0370872, 0.0597026]	94.00 %	2 runs, bitwise identical	measured by us (their artifact)	receipt
RadixArk GLM-5.3-NVFP4 (routed experts NVFP4 e2m1 group 16, rest native)	nvfp4 @4	464.8 GB	0.0510711	[0.0421434, 0.0654185]	93.64 %	2 runs, bitwise identical	measured by us (their artifact)	receipt
incoai GLM-5.3-NVFP4 (routed experts NVFP4 e2m1 group 16, rest native)	nvfp4 @4	464.8 GB	0.0593681	[0.0488554, 0.0770066]	93.29 %	2 runs, bitwise identical	measured by us (their artifact)	receipt
davidsyoung GLM-5.3 EXL3 TR3 3.42bpw (routed experts trellis, TP4 rank-sharded)	exl3-mcg @3.42188	355.2 GB	0.0628419	[0.0510114, 0.0820677]	93.06 %	2 runs, bitwise identical	measured by us (their artifact)	receipt
davidsyoung GLM-5.3 EXL3 TR3 3.25bpw (routed experts trellis, TP4 rank-sharded)	exl3-mcg @3.25	339.4 GB	0.0730595	[0.0597642, 0.0939211]	92.56 %	2 runs, bitwise identical	measured by us (their artifact)	receipt
Inferact GLM-5.3-NVFP4 (routed experts NVFP4 e2m1 group 16, rest native)	nvfp4 @4	464.8 GB	0.0754294	[0.0605579, 0.103742]	92.39 %	2 runs, bitwise identical	measured by us (their artifact)	receipt
davidsyoung GLM-5.3 EXL3 TR3 3.0bpw (routed experts trellis, TP4 rank-sharded)	exl3-mcg @3	316.4 GB	0.0838334	[0.0684587, 0.109309]	92.05 %	2 runs, bitwise identical	measured by us (their artifact)	receipt
drowzeys keys-GLM-5.3-EXL3 (routed experts trellis 3.0 bpw, mcg/mul1)	exl3-trellis @3	330.1 GB	0.102333	[0.0839497, 0.133514]	91.13 %	2 runs, bitwise identical	measured by us (their artifact)	receipt

Disclosures for the rows above (27)

• glm-5.3.fp8-dequantized.corpus5x5-v1 lossy_capture_codec: RECONSTRUCTED, NOT EXECUTED. The candidate was captured from a bf16 materialisation of the stored fp8 weights: every fp8_e4m3 tensor is decoded on the host with its 128x128 weight_scale_inv block scale (engines/tools/layer_outer.py fp8-block-dequant-to-bf16, accumulated fp32, stored bf16) BEFORE it reaches the loader, so no scale can be silently dropped (transformers' plain-cast path would drop all of them). This is the dequantize-and-run methodology: it measures the error of the STORED weights, not of a vendor kernel.
• glm-5.3.fp8-dequantized.corpus5x5-v1 estimator_scope_narrower_than_artifact: WEIGHT-ONLY: expected to understate a served W8A8 deployment; the activation term is not measured. The checkpoint declares activation_scheme: dynamic, so the served model also quantizes activations per token at runtime; that term is absent here. (Wording corrected 2026-09-05: omitting it is expected to understate the served divergence, not a mathematical bound on a mean KL.)
• glm-5.3.fp8-dequantized.corpus5x5-v1 note: Per-window mean 0.022305139008145507, population sd 0.011658841108250139, min 0.0031859275260282391 (final-0012, literary), max 0.066921711801724015 (final-0014, literary) over 25 windows. The macro mean over strata equals the token mean to 1e-16 (every window contributes the same 2,047 positions; the two differ only in fp64 summation order); the token mean is the published value.
• glm-5.3.gguf-unsloth-udq4kxl.corpus5x5-v1 lossy_capture_codec: RECONSTRUCTED, NOT EXECUTED. Every GGUF tensor is dequantized to bf16 on the capture host (gguf-dequant-to-bf16) before the loader, k-quant block traits read from the tensor tables themselves. The decoder is proven BITWISE against gguf-py 0.19.0's own gguf.quants.dequantize on real fetched blocks of this repository at this revision (Q4_K, Q5_K, Q6_K, Q8_0 and the IQ family; engines/tools/gguf-evidence/), so the DECODE is not in question. What is absent is the serving engine: llama.cpp runs these weights through its own kernels and its own KV-cache quantization, and none of that is in this number. This row is advisory because it measures the STORED WEIGHTS, not a llama.cpp deployment.
• glm-5.3.gguf-unsloth-udq4kxl.corpus5x5-v1 quantized_head: HEAD-1d with a QUANTIZED head: this build's lm_head is Q8_0, so the candidate side replayed through its own dequantized head (f5aa1c39b73b...) and the reference through the official bf16 head (864f488a0074...). The head's own quantization error is therefore inside this value -- unlike the three NVFP4 rows, whose heads are the official tensor byte for byte. Read the difference between this row and an NVFP4 row as codec-plus-head, not codec alone.
• glm-5.3.gguf-unsloth-udq4kxl.corpus5x5-v1 note: Per-window mean 0.027546149376942001, population sd 0.014867293502822837, min 0.0022879491144074107 (final-0012, literary), max 0.079814749369237256 (final-0014, literary) over 25 windows. The macro mean over strata equals the token mean to 1e-16 (every window contributes the same 2,047 positions; the two differ only in fp64 summation order); the token mean is the published value.
• glm-5.3.exl3-k4-wrldsuksgo2mars.corpus5x5-v1 lossy_capture_codec: RECONSTRUCTED, NOT EXECUTED. The 57,600 routed-expert trellis payload groups are decoded to bf16 per module on the capture device (exl3-trellis-decode-to-bf16: exllamav3's unpack, tile permutation, two Hadamard GEMMs and su/sv scaling, mcg codebook read from each module's own payload, TF32 pinned off and recorded) and the fp8 tensors the release kept are decoded on the host as for the FP8 row -- all before the loader. Decode evidence: the decoder reproduces engines/tools/exl3hf_surface.py:decode_payload_hf bitwise on real payloads (the suite's own reference decoder, not exllamav3's kernel) and the same path reconstructs a real trellis quant against its bf16 source at the expected K4 error (cosine 0.99773, rel_l2 6.74%). The decode has NOT been proven bitwise against a running exllamav3 kernel, which is why this row is advisory.
• glm-5.3.exl3-k4-wrldsuksgo2mars.corpus5x5-v1 estimator_scope_narrower_than_artifact: The fp8 tensors this release kept carry the source's activation_scheme: dynamic; that runtime term is not measured, so this value is expected to understate a served fp8-activation (W8A8) deployment of it.
• glm-5.3.exl3-k4-wrldsuksgo2mars.corpus5x5-v1 note: Per-window mean 0.044803849964949564, population sd 0.026215181142102181, min 0.0072157360961422135 (final-0012, literary), max 0.14520838316901405 (final-0014, literary) over 25 windows. The macro mean over strata equals the token mean to 1e-16 (every window contributes the same 2,047 positions; the two differ only in fp64 summation order); the token mean is the published value.
• glm-5.3.nvfp4-radixark.corpus5x5-v1 lossy_capture_codec: RECONSTRUCTED, NOT EXECUTED. The 57,600 routed-expert NVFP4 component sets are decoded to bf16 per module before the loader (nvfp4-modelopt-dequant-to-bf16: e2m1 unpack, per-group weight_scale in fp8_e4m3 and the per-tensor weight_scale_2, evaluated in exact fp32 and cast once to bf16). The decoder is proven BITWISE against the ecosystem reference implementation -- compressed-tensors 0.18.0's own unpack_fp4_from_uint8 plus the same LUT/scale math in exact fp32 -- on real range-read tensors of THIS release (max_abs_diff_fp32 exactly 0.0, bitwise after the bf16 cast too; engines/tools/nvfp4-evidence/glm53-nvfp4-parity.json). So the decode is not the reason this row is advisory: the reason is that a weights-only capture runs the STORED weights and not a served NVFP4 kernel.
• glm-5.3.nvfp4-radixark.corpus5x5-v1 activation_quantization_not_captured: WEIGHT-ONLY. The release ships a static per-tensor F32 input_scale beside each of its 57,600 routed-expert modules; a served W4A4 NVFP4 kernel quantizes activations with it and this capture does not. The scales were read and recorded, never applied, so this value is expected to understate a served NVFP4 deployment. It is not a mathematical bound on a mean KL.
• glm-5.3.nvfp4-radixark.corpus5x5-v1 note: Per-window mean 0.051071074118349193, population sd 0.028140927221610379, min 0.010777822592977302 (final-0012, literary), max 0.14531180338616043 (final-0014, literary) over 25 windows. The macro mean over strata equals the token mean to 1e-16 (every window contributes the same 2,047 positions; the two differ only in fp64 summation order); the token mean is the published value.
• glm-5.3.nvfp4-incoai.corpus5x5-v1 lossy_capture_codec: RECONSTRUCTED, NOT EXECUTED. The 57,600 routed-expert NVFP4 component sets are decoded to bf16 per module before the loader (nvfp4-modelopt-dequant-to-bf16: e2m1 unpack, per-group weight_scale in fp8_e4m3 and the per-tensor weight_scale_2, evaluated in exact fp32 and cast once to bf16). The decoder is proven BITWISE against the ecosystem reference implementation -- compressed-tensors 0.18.0's own unpack_fp4_from_uint8 plus the same LUT/scale math in exact fp32 -- on real range-read tensors of THIS release (max_abs_diff_fp32 exactly 0.0, bitwise after the bf16 cast too; engines/tools/nvfp4-evidence/glm53-nvfp4-parity.json). So the decode is not the reason this row is advisory: the reason is that a weights-only capture runs the STORED weights and not a served NVFP4 kernel.
• glm-5.3.nvfp4-incoai.corpus5x5-v1 activation_quantization_not_captured: WEIGHT-ONLY. The release ships a static per-tensor F32 input_scale beside each of its 57,600 routed-expert modules; a served W4A4 NVFP4 kernel quantizes activations with it and this capture does not. The scales were read and recorded, never applied, so this value is expected to understate a served NVFP4 deployment. It is not a mathematical bound on a mean KL.
• glm-5.3.nvfp4-incoai.corpus5x5-v1 note: Per-window mean 0.0593681245487735, population sd 0.033459168535851555, min 0.0098494657264303568 (final-0012, literary), max 0.17619923954878838 (final-0014, literary) over 25 windows. The macro mean over strata equals the token mean to 1e-16 (every window contributes the same 2,047 positions; the two differ only in fp64 summation order); the token mean is the published value.
• glm-5.3.exl3-tr3-3.42bpw-davidsyoung.corpus5x5-v1 lossy_capture_codec: RECONSTRUCTED, NOT EXECUTED. Every routed-expert trellis payload group is decoded to bf16 per module on the capture device (exl3-trellis-decode-to-bf16, TP4 rank shards composed per module) before the loader; the decoder reproduces engines/tools/exl3hf_surface.py:decode_payload_hf bitwise on real payloads (the suite's own reference decoder, not exllamav3's kernel) and reconstructs a real trellis quant against its bf16 source at the expected error. It has NOT been proven bitwise against a running exllamav3 kernel, which is why this row is advisory.
• glm-5.3.exl3-tr3-3.42bpw-davidsyoung.corpus5x5-v1 note: Per-window mean 0.062841891548989365, population sd 0.037338302064760603, min 0.007424164748414566 (final-0012, literary), max 0.19156791191512221 (final-0014, literary) over 25 windows. The macro mean over strata equals the token mean to 1e-16 (every window contributes the same 2,047 positions; the two differ only in fp64 summation order); the token mean is the published value.
• glm-5.3.exl3-tr3-3.25bpw-davidsyoung.corpus5x5-v1 lossy_capture_codec: RECONSTRUCTED, NOT EXECUTED. Every routed-expert trellis payload group is decoded to bf16 per module on the capture device (exl3-trellis-decode-to-bf16, TP4 rank shards composed per module) before the loader; the decoder reproduces engines/tools/exl3hf_surface.py:decode_payload_hf bitwise on real payloads (the suite's own reference decoder, not exllamav3's kernel) and reconstructs a real trellis quant against its bf16 source at the expected error. It has NOT been proven bitwise against a running exllamav3 kernel, which is why this row is advisory.
• glm-5.3.exl3-tr3-3.25bpw-davidsyoung.corpus5x5-v1 note: Per-window mean 0.073059477496064701, population sd 0.04149044465864947, min 0.01180684193920154 (final-0012, literary), max 0.20926743181562468 (final-0014, literary) over 25 windows. The macro mean over strata equals the token mean to 1e-16 (every window contributes the same 2,047 positions; the two differ only in fp64 summation order); the token mean is the published value.
• glm-5.3.nvfp4-inferact.corpus5x5-v1 lossy_capture_codec: RECONSTRUCTED, NOT EXECUTED. The 57,600 routed-expert NVFP4 component sets are decoded to bf16 per module before the loader (nvfp4-modelopt-dequant-to-bf16: e2m1 unpack, per-group weight_scale in fp8_e4m3 and the per-tensor weight_scale_2, evaluated in exact fp32 and cast once to bf16). The decoder is proven BITWISE against the ecosystem reference implementation -- compressed-tensors 0.18.0's own unpack_fp4_from_uint8 plus the same LUT/scale math in exact fp32 -- on real range-read tensors of THIS release (max_abs_diff_fp32 exactly 0.0, bitwise after the bf16 cast too; engines/tools/nvfp4-evidence/glm53-nvfp4-parity.json). So the decode is not the reason this row is advisory: the reason is that a weights-only capture runs the STORED weights and not a served NVFP4 kernel.
• glm-5.3.nvfp4-inferact.corpus5x5-v1 activation_quantization_not_captured: WEIGHT-ONLY. The release ships a static per-tensor F32 input_scale beside each of its 57,600 routed-expert modules; a served W4A4 NVFP4 kernel quantizes activations with it and this capture does not. The scales were read and recorded, never applied, so this value is expected to understate a served NVFP4 deployment. It is not a mathematical bound on a mean KL.
• glm-5.3.nvfp4-inferact.corpus5x5-v1 note: Per-window mean 0.075429362164713742, population sd 0.05057088279896952, min 0.0089240014345735898 (final-0012, literary), max 0.26522688023444346 (final-0014, literary) over 25 windows. The macro mean over strata equals the token mean to 1e-16 (every window contributes the same 2,047 positions; the two differ only in fp64 summation order); the token mean is the published value.
• glm-5.3.exl3-tr3-3.0bpw-davidsyoung.corpus5x5-v1 lossy_capture_codec: RECONSTRUCTED, NOT EXECUTED. Every routed-expert trellis payload group is decoded to bf16 per module on the capture device (exl3-trellis-decode-to-bf16, TP4 rank shards composed per module) before the loader; the decoder reproduces engines/tools/exl3hf_surface.py:decode_payload_hf bitwise on real payloads (the suite's own reference decoder, not exllamav3's kernel) and reconstructs a real trellis quant against its bf16 source at the expected error. It has NOT been proven bitwise against a running exllamav3 kernel, which is why this row is advisory.
• glm-5.3.exl3-tr3-3.0bpw-davidsyoung.corpus5x5-v1 note: Per-window mean 0.083833394938045827, population sd 0.0490654872424078, min 0.011591449983713651 (final-0012, literary), max 0.25238779656109533 (final-0014, literary) over 25 windows. The macro mean over strata equals the token mean to 1e-16 (every window contributes the same 2,047 positions; the two differ only in fp64 summation order); the token mean is the published value.
• glm-5.3.exl3-keys-drowzeys.corpus5x5-v1 lossy_capture_codec: RECONSTRUCTED, NOT EXECUTED. Every routed-expert trellis payload group is decoded to bf16 per module on the capture device (exl3-trellis-decode-to-bf16, mcg on layer 3 and mul1 on layers 4-77, each read from the module's own payload) before the loader; the decoder reproduces engines/tools/exl3hf_surface.py:decode_payload_hf bitwise on real payloads (the suite's own reference decoder, not exllamav3's kernel) and reconstructs a real trellis quant against its bf16 source at the expected error. It has NOT been proven bitwise against a running exllamav3 kernel, which is why this row is advisory.
• glm-5.3.exl3-keys-drowzeys.corpus5x5-v1 record_note: NON-ROUTED PATH IS FP8-DERIVED. This artifact's attention, dense-MLP and shared-expert tensors are the FP8 release's block-dequantized weights stored at fp16 (byte evidence engines/tools/layer-outer-evidence/drowzeys-nonrouted-provenance.json; scope attn./mlp./moe.shared_expert = quantized:fp8_e4m3@8), while davidsyoung's three releases carry the BF16 release's values. The 0.0185-nat gap between this row and measurement--glm-5.3.exl3-tr3-3.0bpw-davidsyoung.corpus5x5-v1 (this row higher on 25 of 25 windows) therefore mixes two effects -- the codec on the routed experts and the FP8 release's non-expert error, itself 0.0223 nats on the FP8 row -- and is NOT a clean codec-vs-codec comparison at 3.0 bpw. Corrected 2026-09-05; until then the row's artifact record called the non-routed path native.
• glm-5.3.exl3-keys-drowzeys.corpus5x5-v1 note: Per-window mean 0.10233258694757998, population sd 0.059543079503892628, min 0.01697183535100235 (final-0012, literary), max 0.30580890039836339 (final-0014, literary) over 25 windows. The macro mean over strata equals the token mean to 1e-16 (every window contributes the same 2,047 positions; the two differ only in fp64 summation order); the token mean is the published value.

GLM-5.3-Flash

model--zai-org.glm-5.3-flash -- published by Z.ai. Tokenizer glm-5.3-flash, vocabulary 154880.

Panel: malaiwah GLM-5.3-Flash distribution-fidelity suite v5 -- 5,120 contexts

Panel disclosure -- no_known_deviations: No deviation from this registry's default protocol is known for this record.

Group cmp--9b009314102d9e8b -- 1 row

Panel panel--glm53.malaiwah.suite-v5-10m -- malaiwah GLM-5.3-Flash distribution-fidelity suite v5 -- 5,120 contexts 5120 contexts x 2047 scored positions = 10,480,640 scored positions, score_from 0 sealed: yes (token digest 2e0ea09683564554...) -- contamination scan: yes, 0 hits Reference (teacher) reference--malaiwah.glm53-bf16-vllm.suite-v5-10m -- native_bf16, artifact artifact--zai-org.glm-5.3-flash-bf16.b1967181 @b1967181a3917ae70a437f4884748f6b8e3a1f4d Metric mean_tokenwise_kld, direction reference_to_candidate, accumulation float64 Estimation surface stack_relation same_stack, head_policy shared_reference_head Comparability key cmp--9b009314102d9e8b Like-for-like predicate comparable: unknown -- no recorded difference, but harness, replay_backend, replay_env, stack are unrecorded for at least one member, so homogeneity cannot be certified. Machine-readable form with per-dimension values: this key's comparability block in index.json.

What this table is. Every row here shares the comparability key above: the same tokens, the same teacher capture, the same metric and direction, the same estimator precision, the same stack relation and the same head policy. That makes them CANDIDATES for ranking -- the key is a necessary partition, not a certificate. Whether they are also like-for-like on the dimensions the key omits (lane, pipeline, scope coverage, hardware) is what the predicate line above answers.

Rank is not a verdict. The table is sorted by fidelity alone, and fidelity buys bits: a larger, higher-bitrate quant will usually sit above a smaller one, which is not news. Read the Size and Codec columns before reading the order, and compare like against like.

What it is NOT comparable to. The nearest neighbouring groups differ in:

• cmp--e6cdd07242bdde05 (1 row): panel_id panel--glm53.malaiwah.suite-v5-10m -> panel--glm53.malaiwah.suite-v5-10m.scorefrom1024; reference_id reference--malaiwah.glm53-bf16-vllm.suite-v5-10m -> reference--malaiwah.glm53-bf16-vllm.suite-v5-10m.scorefrom1024

Those numbers are in this file, under their own headings. Quoting one under the other heading is the mistake this layout exists to prevent: the key is a function of the panel, the teacher, the metric, the direction and the estimator, and the validator recomputes it from those fields rather than trusting the stamped value. What that catches is a row filed under a key its own fields do not produce. It does not catch a number attributed to the wrong panel in the first place -- no offline checker can. That is what the receipt digests on every row are for.

Also, and always: every table for a different model. A KL number is a divergence over one model's own vocabulary against that model's own teacher. It is not a quality score that can be carried between models.

Single-row group. This number has nothing in the registry to be ranked against. It is a stated fact, not a placing.

Artifact	Codec	Size	mean_tokenwise_kld (nats)	CI95	Top-1	Runs	Attribution	Receipt
GLM-5.3-Flash official FP8	fp8_e4m3 @8	328.4 GB	0.0281039	[0.0272053, 0.0289822]	94.27 %	1 run, unevidenced	measured by us (their artifact)	receipt

The same artifact, measured elsewhere in this file. One of the artifacts below also carries a number in another table -- on a different panel, teacher or estimator -- and the widest of those spans 51%. None of the readings is wrong and none is interchangeable with another. Quoting one of them as the number for the artifact, without its table, is the misuse this registry exists to make obvious.

• GLM-5.3-Flash official FP8 -- 4 values here, from 0.0186653 to 0.0281039 nats (51% apart). Other tables: cmp--4a8630bdcadab97f, cmp--e6cdd07242bdde05, cmp--eee09298c558ab21.

Disclosures for the rows above (1)

• glm53.official-fp8.malaiwah-suite-v5-10m single_run: One pass; determinism not established for this row.

Panel: malaiwah GLM-5.3-Flash suite v5, scored from position 1024

Derived from panel--glm53.malaiwah.suite-v5-10m by scoring_window_change: score_from 0 -> 1024. Identical tokens, half the scored positions, and a materially different number: 0.028104 becomes 0.018794 on the same artifact and the same teacher. This is the clearest demonstration in the registry that the scored-position policy is part of panel identity.

Panel disclosure -- no_known_deviations: No deviation from this registry's default protocol is known for this record.

Group cmp--e6cdd07242bdde05 -- 1 row

Panel panel--glm53.malaiwah.suite-v5-10m.scorefrom1024 -- malaiwah GLM-5.3-Flash suite v5, scored from position 1024 5120 contexts x 1023 scored positions = 5,237,760 scored positions, score_from 1024, windowed sealed: yes (token digest 2e0ea09683564554...) -- contamination scan: yes, 0 hits Reference (teacher) reference--malaiwah.glm53-bf16-vllm.suite-v5-10m.scorefrom1024 -- native_bf16, artifact artifact--zai-org.glm-5.3-flash-bf16.b1967181 @b1967181a3917ae70a437f4884748f6b8e3a1f4d Metric mean_tokenwise_kld, direction reference_to_candidate, accumulation float64 Estimation surface stack_relation same_stack, head_policy shared_reference_head Comparability key cmp--e6cdd07242bdde05 Like-for-like predicate comparable: unknown -- no recorded difference, but harness, replay_backend, replay_env, stack are unrecorded for at least one member, so homogeneity cannot be certified. Machine-readable form with per-dimension values: this key's comparability block in index.json.

What this table is. Every row here shares the comparability key above: the same tokens, the same teacher capture, the same metric and direction, the same estimator precision, the same stack relation and the same head policy. That makes them CANDIDATES for ranking -- the key is a necessary partition, not a certificate. Whether they are also like-for-like on the dimensions the key omits (lane, pipeline, scope coverage, hardware) is what the predicate line above answers.

Rank is not a verdict. The table is sorted by fidelity alone, and fidelity buys bits: a larger, higher-bitrate quant will usually sit above a smaller one, which is not news. Read the Size and Codec columns before reading the order, and compare like against like.

What it is NOT comparable to. The nearest neighbouring groups differ in:

• cmp--9b009314102d9e8b (1 row): panel_id panel--glm53.malaiwah.suite-v5-10m.scorefrom1024 -> panel--glm53.malaiwah.suite-v5-10m; reference_id reference--malaiwah.glm53-bf16-vllm.suite-v5-10m.scorefrom1024 -> reference--malaiwah.glm53-bf16-vllm.suite-v5-10m

Those numbers are in this file, under their own headings. Quoting one under the other heading is the mistake this layout exists to prevent: the key is a function of the panel, the teacher, the metric, the direction and the estimator, and the validator recomputes it from those fields rather than trusting the stamped value. What that catches is a row filed under a key its own fields do not produce. It does not catch a number attributed to the wrong panel in the first place -- no offline checker can. That is what the receipt digests on every row are for.

Also, and always: every table for a different model. A KL number is a divergence over one model's own vocabulary against that model's own teacher. It is not a quality score that can be carried between models.

Single-row group. This number has nothing in the registry to be ranked against. It is a stated fact, not a placing.

Artifact	Codec	Size	mean_tokenwise_kld (nats)	CI95	Top-1	Runs	Attribution	Receipt
GLM-5.3-Flash official FP8	fp8_e4m3 @8	328.4 GB	0.0187943	[0.0180739, 0.0194941]	95.12 %	1 run, unevidenced	measured by us (their artifact)	receipt

The same artifact, measured elsewhere in this file. One of the artifacts below also carries a number in another table -- on a different panel, teacher or estimator -- and the widest of those spans 51%. None of the readings is wrong and none is interchangeable with another. Quoting one of them as the number for the artifact, without its table, is the misuse this registry exists to make obvious.

• GLM-5.3-Flash official FP8 -- 4 values here, from 0.0186653 to 0.0281039 nats (51% apart). Other tables: cmp--4a8630bdcadab97f, cmp--9b009314102d9e8b, cmp--eee09298c558ab21.

Disclosures for the rows above (2)

• glm53.official-fp8.malaiwah-suite-v5-10m.scorefrom1024 single_run: One pass; determinism not established.
• glm53.official-fp8.malaiwah-suite-v5-10m.scorefrom1024 note: Same tokens, same artifact, same teacher as the 0.028104 row. Dropping the first 1024 scored positions of every context moves the number by 33%. That is why the scored-position policy is part of panel identity.

Panel: brandonmusic GLM-5.3-Flash sealed qualification panel v1 -- 25 final windows

Panel disclosure -- weak_contamination_guard: This panel's only contamination guard is ROLE SEPARATION: the 25 'final' windows are drawn from the same packed corpus as the 384 fit / 128 conditional-fit / 64 selection / 64 confirmation windows and are declared qualification-only. No lexical or n-gram scan is published, and the underlying document provenance is published only as a digest. This is materially weaker than the malaiwah v5 suites, which run a 12-word shingle whole-document pre-exclusion and report 0 hits. Do not describe the two guards as equivalent. It applies equally to every row on this panel, so it does not disturb comparisons WITHIN the panel.

This panel carries 3 separate comparability groups. They are different measurements of different things and are never merged.

Group cmp--202b717f3219c414 -- 11 rows

Panel panel--glm53.brandonmusic.final25 -- brandonmusic GLM-5.3-Flash sealed qualification panel v1 -- 25 final windows 25 contexts x 2047 scored positions = 51,175 scored positions, score_from 0 sealed: yes (token digest 6bafe3283c54bc93...) -- contamination scan: NOT RUN Reference (teacher) reference--brandonmusic.glm53-bf16-fp32-logits.final25 -- native_bf16, artifact artifact--zai-org.glm-5.3-flash-bf16.a6c167b6 @a6c167b62691b2bac901344b65cb651a70f53e43 Metric mean_of_run_means_tokenwise_kld, direction reference_to_candidate, accumulation float64 Estimation surface stack_relation same_stack, head_policy native_head Comparability key cmp--202b717f3219c414 Like-for-like predicate comparable: false -- a RECORDED secondary dimension differs across members: lane, pipeline, scope. Equal keys make these rows candidates for comparison, not certified like-for-like; ranking across the differing dimension attributes a lane/pipeline/hardware/scope effect to quantization quality. Machine-readable form with per-dimension values: this key's comparability block in index.json.

What this table is. Every row here shares the comparability key above: the same tokens, the same teacher capture, the same metric and direction, the same estimator precision, the same stack relation and the same head policy. That makes them CANDIDATES for ranking -- the key is a necessary partition, not a certificate. Whether they are also like-for-like on the dimensions the key omits (lane, pipeline, scope coverage, hardware) is what the predicate line above answers.

Rank is not a verdict. The table is sorted by fidelity alone, and fidelity buys bits: a larger, higher-bitrate quant will usually sit above a smaller one, which is not news. Read the Size and Codec columns before reading the order, and compare like against like.

What it is NOT comparable to. The nearest neighbouring groups differ in:

• cmp--f0823827adb15376 (2 rows): reference_id reference--brandonmusic.glm53-bf16-fp32-logits.final25 -> reference--malaiwah.glm53-bf16-hf.brandonmusic-final25; metric_name mean_of_run_means_tokenwise_kld -> mean_tokenwise_kld
• cmp--4a8630bdcadab97f (2 rows): metric_name mean_of_run_means_tokenwise_kld -> mean_tokenwise_kld; stack_relation same_stack -> cross_stack
• cmp--2b9c401d13806d7e (4 rows): panel_id panel--glm53.brandonmusic.final25 -> panel--glm53.brandonmusic.final25-clean17; reference_id reference--brandonmusic.glm53-bf16-fp32-logits.final25 -> reference--brandonmusic.glm53-bf16-fp32-logits.final25-clean17

Those numbers are in this file, under their own headings. Quoting one under the other heading is the mistake this layout exists to prevent: the key is a function of the panel, the teacher, the metric, the direction and the estimator, and the validator recomputes it from those fields rather than trusting the stamped value. What that catches is a row filed under a key its own fields do not produce. It does not catch a number attributed to the wrong panel in the first place -- no offline checker can. That is what the receipt digests on every row are for.

Also, and always: every table for a different model. A KL number is a divergence over one model's own vocabulary against that model's own teacher. It is not a quality score that can be carried between models.

8 of this group's 11 rows came off a different measurement lane (streaming) and are tabled on their own below, not mixed into the ordering here. The key does not carry the lane; this file does.

Artifact	Codec	Size	mean_of_run_means_tokenwise_kld (nats)	CI95	Top-1	Runs	Attribution	Receipt
malaiwah GLM-5.3-Flash TR3 6bpw (K6)	exl3-mcg @6	253.5 GB	0.0137234	[0.0111548, 0.0166267]	--	5 runs, bitwise identical	measured by us	receipt
brandonmusic GLM-5.3-Flash tr3 4bpw	exl3-mcg @4	175.6 GB	0.0245546	[0.019433, 0.035881]	--	5 runs, bitwise identical	reported by brandonmusic	receipt
0xSero GLM-5.3-Flash EXL3 Q4 (Dione, TP4-sliced)	exl3-mcg @4	187.6 GB	0.0272628	--	--	5 runs, bitwise identical	measured by us (their artifact)	receipt

Lane streaming -- 8 of this group's 11 rows

A different lane. Same key, and that is exactly the problem this table solves. The comparability key is a function of the panel, the teacher, the metric, the direction, the estimator precision, the stack relation and the head policy -- and these rows match the table above on all seven. What they do not share is the machine and the code path that produced the candidate logits, and lanes are not interchangeable. Sorting them into one list would read as a ranking; where the same artifact appears in both, it is one set of weights measured twice, not two quants.

What the lane is (from pipeline--malaiwah.glm53-stream-packed-kld): 1 device, expert-parallel width 8 emulated in one process, routed-expert combine order fp32.

Bridge to the sealed-ep8 lane, measured on this panel: signed delta -8.4958e-06 nats on the mean against measurement--glm53.k6-6bpw.brandonmusic-final25, worst single window 0.00028735, over 25 windows. Tokenwise KL array matches the sealed run: no. The runner's own verdict on whether this may be published as a reproduction of the sealed number: no (verdict LARGER_DELTA_SEE_DISCLOSURE).

That bridge is one artifact's, on one panel. It is not a constant and it is not subtractable: a row in this table whose artifact has no sealed-lane row has no measured offset at all, and says so in its own bias line.

Artifact	Codec	Size	mean_of_run_means_tokenwise_kld (nats)	Excess over control (nats)	CI95	Top-1	Runs	Attribution	Receipt
GLM-5.3-Flash BF16 @a6c167b6 (measurement floor)	bf16	--	0.0115059	--	--	--	2 runs, bitwise identical	measured by us (their artifact)	receipt
malaiwah GLM-5.3-Flash TR3 8bpw (K8)	exl3-mcg @8	331.4 GB	0.0123842	0.000878268	[0.00999057, 0.0152193]	--	2 runs, bitwise identical	measured by us	receipt
malaiwah GLM-5.3-Flash TR3 6bpw (K6)	exl3-mcg @6	253.5 GB	0.0137149	0.00220897	[0.0111567, 0.01663]	96.56 %	2 runs, bitwise identical	measured by us	receipt
Mia-AiLab GLM-5.3-Flash EXL3 TR3 4bpw (byte-identical mirror of brandonmusic's)	exl3-mcg @4	175.7 GB	0.0255034	0.0139975	--	95.31 %	2 runs, bitwise identical	measured by us (their artifact)	receipt
turboderp GLM-5.3-Flash EXL3 4.05bpw (stock exllamav3, mul1, quantized head)	exl3-mul1 @4.05	165.2 GB	0.0255264	0.0140205	--	95.10 %	2 runs, bitwise identical	measured by us (their artifact)	receipt
0xSero GLM-5.3-Flash EXL3 3.0bpw (Dione, K3, TP4-sliced, native BF16 head)	exl3-mcg @3	149.6 GB	0.0505012	0.0389953	--	93.00 %	2 runs, bitwise identical	measured by us (their artifact)	receipt
turboderp GLM-5.3-Flash EXL3 2.05bpw (stock exllamav3, mul1, quantized head at 5 bits)	exl3-mul1 @2.05	85.2 GB	0.121638	0.110132	--	88.92 %	2 runs, bitwise identical	measured by us (their artifact)	receipt
vcruz305 GLM-5.3-Flash EXL3 K2 (stock-exllamav3 HF layout, mcg, routed experts only, native BF16 head)	exl3-mcg @2	97.8 GB	0.15521	0.143704	--	87.27 %	2 runs, bitwise identical	measured by us (their artifact)	receipt

Excess over control (nats) = this row's value minus its named floor's value (comparability.bias.floor_measurement_ref) -- the raw number with this lane's own measurement floor netted out. Until 2026-08-31 this column was named Attributable (nats); the rename is peer-review P1-05, and it is a claim change, not a cosmetic one: the difference D(P||Q_quant) - D(P||Q_control) is not itself a divergence, can be negative, and isolates quantization only if the two paths differ by nothing else -- an assumption this project's own pipeline and hardware studies show is non-trivial. It is an estimate, not an identity: KL is not additive, and the subtraction is only meaningful because both terms are small and share the same reference and the same lane. Do not quote a RATIO of two of these numbers without uncertainty: a ratio of small residuals magnifies control error. A row with no floor named shows --, not zero: absence of a floor is not evidence the floor is zero. BIAS-002/004/006 guarantee any floor named here shares this row's comparability key, measures unquantized weights, and was measured on this row's own lane -- so this column can never mix a floor from a different panel, a different kind of thing, or a different lane into the subtraction.

Bias on GLM-5.3-Flash BF16 @a6c167b6 -- other, direction unknown. THIS ROW IS THE FLOOR for the 'streaming' lane: it replays the reference's own unquantized weights through the SAME streaming harness that scored every other row on this pipeline, so its divergence against the stored teacher logits is the lane's zero-point, not a quantization result. It is NOT the cross-stack floor recorded elsewhere in this registry (a different pipeline, a different lane, a different comparability key) and is never interchangeable with it: subtracting one lane's floor from another lane's row is exactly the mistake BIAS-006 exists to catch. The lane's offset against the sealed-ep8 lane is NOT measured for this artifact: no sealed-lane counterpart to this profile exists to bridge against.

Bias on malaiwah GLM-5.3-Flash TR3 8bpw (K8) -- other, direction unknown. Measured on the 'streaming' lane, whose offset against the sealed-ep8 lane is known to be non-zero but was NOT measured for this artifact: no sealed-lane row for it exists to bridge against. The lane offset measured for a sibling artifact on this panel is not transferable -- it is a property of the routing, not a constant. This lane's own measurement floor (measurement--glm53.bf16-stream-floor.brandonmusic-final25) is 0.011505922619330299 nats; netting it out gives an estimated excess_over_control of 0.0008782684041065674 nats here (called 'quantization-attributable error' before 2026-08-31, renamed per peer-review P1-05: the difference estimates excess divergence over the same-lane unquantized control and is not a causal attribution) -- an estimate, not an identity, because KL is not additive, and it is only meaningful because both terms are small and share the same reference and lane.

Bias on malaiwah GLM-5.3-Flash TR3 6bpw (K6) -- other, direction downward. Lane offset, MEASURED not estimated: this 'streaming'-lane run scores 0.013714888822596553 against the sealed-ep8 lane's 0.013723384665701147 on the same panel, a signed delta of -8.495843104593809e-06 nats (|max| 0.00028735280093581186 on any one of 25 windows). The tokenwise KL array does NOT match the sealed one, and the runner's own verdict is publishable_as_reproduction=False, so this number stands beside the sealed one rather than replacing it. This lane's own measurement floor (measurement--glm53.bf16-stream-floor.brandonmusic-final25) is 0.011505922619330299 nats; netting it out gives an estimated excess_over_control of 0.0022089662032662542 nats here (called 'quantization-attributable error' before 2026-08-31, renamed per peer-review P1-05: the difference estimates excess divergence over the same-lane unquantized control and is not a causal attribution) -- an estimate, not an identity, because KL is not additive, and it is only meaningful because both terms are small and share the same reference and lane.

Bias on Mia-AiLab GLM-5.3-Flash EXL3 TR3 4bpw (byte-identical mirror of brandonmusic's) -- other, direction unknown. Measured on the 'streaming' lane. Unlike every other streaming row, this artifact HAS a sealed-lane sibling to bridge against: the same bytes read 0.024554564249958208 there (author-reported, brandonmusic's own five-run receipt on his own stack), so the streaming-lane number sits +0.000948863 nats from it -- a LANE-PLUS-STACK offset, not a lane offset, because the reader digests differ too (1fb3be87... vs 1ccce446...). This lane's own measurement floor (measurement--glm53.bf16-stream-floor.brandonmusic-final25) is 0.011505922619330299 nats; netting it out gives an estimated excess_over_control of 0.01399750501503347 nats here (called 'quantization-attributable error' before 2026-08-31, renamed per peer-review P1-05: the difference estimates excess divergence over the same-lane unquantized control and is not a causal attribution) -- an estimate, not an identity, because KL is not additive, and it is only meaningful because both terms are small and share the same reference and lane.

Bias on turboderp GLM-5.3-Flash EXL3 4.05bpw (stock exllamav3, mul1, quantized head) -- other, direction unknown. Measured on the 'streaming' lane, whose offset against the sealed-ep8 lane is known to be non-zero but was NOT measured for this artifact: no sealed-lane row for it exists to bridge against. This lane's own measurement floor (measurement--glm53.bf16-stream-floor.brandonmusic-final25) is 0.011505922619330299 nats; netting it out gives an estimated excess_over_control of 0.014020504296142185 nats here (called 'quantization-attributable error' before 2026-08-31, renamed per peer-review P1-05: the difference estimates excess divergence over the same-lane unquantized control and is not a causal attribution) -- an estimate, not an identity, because KL is not additive, and it is only meaningful because both terms are small and share the same reference and lane.

Bias on 0xSero GLM-5.3-Flash EXL3 3.0bpw (Dione, K3, TP4-sliced, native BF16 head) -- other, direction unknown. Measured on the 'streaming' lane, whose offset against the sealed-ep8 lane is known to be non-zero and is NOT measured for this artifact: it has no sealed-lane row to bridge against. Its 4-bpw SIBLING does (measurement--glm53.dione-q4.brandonmusic-final25, 0.027262784814670614 on the sealed lane), but a lane offset is a property of the routing, not a constant, so it does not transfer between rungs of a ladder. This lane's own measurement floor (measurement--glm53.bf16-stream-floor.brandonmusic-final25) is 0.011505922619330299 nats; netting it out gives an estimated excess_over_control of 0.03899531884609326 nats here (called 'quantization-attributable error' before 2026-08-31, renamed per peer-review P1-05: the difference estimates excess divergence over the same-lane unquantized control and is not a causal attribution) -- an estimate, not an identity, because KL is not additive, and it is only meaningful because both terms are small and share the same reference and lane.

Bias on turboderp GLM-5.3-Flash EXL3 2.05bpw (stock exllamav3, mul1, quantized head at 5 bits) -- other, direction unknown. Measured on the 'streaming' lane, whose offset against the sealed-ep8 lane is known to be non-zero but was NOT measured for this artifact. This lane's own measurement floor (measurement--glm53.bf16-stream-floor.brandonmusic-final25) is 0.011505922619330299 nats; netting it out gives an estimated excess_over_control of 0.11013175411406427 nats here (called 'quantization-attributable error' before 2026-08-31, renamed per peer-review P1-05: the difference estimates excess divergence over the same-lane unquantized control and is not a causal attribution) -- an estimate, not an identity, because KL is not additive, and it is only meaningful because both terms are small and share the same reference and lane.

Bias on vcruz305 GLM-5.3-Flash EXL3 K2 (stock-exllamav3 HF layout, mcg, routed experts only, native BF16 head) -- other, direction unknown. Measured on the 'streaming' lane, whose offset against the sealed-ep8 lane is known to be non-zero and is NOT measured for this artifact: it has no sealed-lane row to bridge against, and no sibling of its own on either lane. This lane's own measurement floor (measurement--glm53.bf16-stream-floor.brandonmusic-final25) is 0.011505922619330299 nats; netting it out gives an estimated excess_over_control of 0.14370363229489977 nats here (called 'quantization-attributable error' before 2026-08-31, renamed per peer-review P1-05: the difference estimates excess divergence over the same-lane unquantized control and is not a causal attribution) -- an estimate, not an identity, because KL is not additive, and it is only meaningful because both terms are small and share the same reference and lane.

The same artifact, measured elsewhere in this file. 4 of the artifacts below also carry a number in another table -- on a different panel, teacher or estimator -- and the widest of those spans 18%. None of the readings is wrong and none is interchangeable with another. Quoting one of them as the number for the artifact, without its table, is the misuse this registry exists to make obvious.

the 4 artifacts and their ranges

• GLM-5.3-Flash BF16 @a6c167b6 -- 4 values here, from 0 to 0.0127116 nats (0% apart). Other tables: cmp--4a8630bdcadab97f, cmp--eee09298c558ab21, cmp--f0823827adb15376.
• brandonmusic GLM-5.3-Flash tr3 4bpw -- 3 values here, from 0.0227508 to 0.0249488 nats (10% apart). Other tables: cmp--18990ab191ea7a67, cmp--2b9c401d13806d7e.
• malaiwah GLM-5.3-Flash TR3 6bpw (K6) -- 6 values here, from 0.011676 to 0.0137234 nats (18% apart). Other tables: cmp--2b9c401d13806d7e.
• malaiwah GLM-5.3-Flash TR3 8bpw (K8) -- 2 values here, from 0.0108294 to 0.0123842 nats (14% apart). Other tables: cmp--2b9c401d13806d7e.

Disclosures for the rows above (37)

• glm53.bf16-stream-floor.brandonmusic-final25 reduced_run_count: cold_run_deviation (verbatim from the receipt): 2 cold runs, not 5 (budget; disclosed)
• glm53.bf16-stream-floor.brandonmusic-final25 non_sealed_lane: Produced by the 'streaming' lane, not the sealed-ep8 lane. The lane's offset against the sealed lane is NOT measured for this artifact: no sealed-lane row for it exists to bridge against. This row is itself the streaming lane's measurement floor -- the zero-point the K6-stream and K8-stream rows in this same table subtract to obtain their own excess_over_control (formerly: quantization-attributable error; P1-05) (see their bias blocks).
• glm53.bf16-stream-floor.brandonmusic-final25 note: CONTROL ROW / STREAMING-LANE MEASUREMENT FLOOR. Not the cross-stack floor (measurement--glm53.bf16-replay-floor.brandonmusic-final25, 0.012712 nats, pipeline--malaiwah.glm53-crosscheck): a different pipeline, a different lane, a different comparability key -- BIAS-002 already keeps the two apart by key, and BIAS-006 additionally forbids naming one as the other's floor even inside a shared key. Provenance of the fields the summary receipt does not carry: metric.direction and estimator.accumulation_dtype are SUPPLIED as reference_to_candidate / float64, matching every other row on this pipeline, because the scorer is the same unmodified tools/k6_kld_report.py. measurement_scope.scored_positions and contexts are SUPPLIED as the panel's own 51,175 positions over 25 contexts (25 x 2047) -- like the K8-stream row, no verdict receipt exists for this profile to read the window count from. determinism.identical_across_runs is RECOMPUTED from run_means and distinct_tokenwise_kld_sha256; the receipt's own bitwise_deterministic flag was checked against that, not copied. cold_run_count (2) was checked against len(run_means) and len(kld_report_sha256), both 2. No clean17 sibling: receipt registry/receipts/malaiwah/stream-bf16-kld.json is scalar-only (run_means + a tokenwise digest, no per_window block), so the calibration-clean scope cannot be recomputed without re-running the measurement.
• glm53.k8-8bpw-stream.brandonmusic-final25 reduced_run_count: cold_run_deviation (verbatim from the receipt): 2 cold runs, not 5 (budget; disclosed)
• glm53.k8-8bpw-stream.brandonmusic-final25 non_sealed_lane: Produced by the 'streaming' lane, not the sealed-ep8 lane. The lane's offset against the sealed lane is NOT measured for this artifact: no sealed-lane row for it exists to bridge against.
• glm53.k8-8bpw-stream.brandonmusic-final25 note: This receipt does not name its lane. Its schema string is malaiwah.glm53-k8-packed-kld-summary.v1 and its profile reads 'k8-tp4' -- neither carries the '-stream-' marker the K6 summary's family name does -- so 'streaming' here is OPERATOR-ASSERTED (operator inventory, 2026-08-28) and not read off the file. It is recorded as the more caveated of the two possibilities on purpose: if the assertion is wrong the row is under-claimed, never over-claimed. Also supplied rather than read: metric.direction, estimator.accumulation_dtype, measurement_scope.scored_positions and contexts -- this family is a scalar summary and states none of them, and unlike the K6 row there is no verdict receipt here to read the window count from. No top-1 agreement was produced for this run. determinism.identical_across_runs is RECOMPUTED from run_means and distinct_tokenwise_kld_sha256. comparability.bias.floor_measurement_ref: SUPPLIED by --floor-measurement once the streaming-lane floor row below existed; build_row checked it was measured on this SAME lane before writing the reference (exit 7 otherwise).
• glm53.k6-6bpw-stream.brandonmusic-final25 reduced_run_count: cold_run_deviation (verbatim from the receipt): 2 cold runs, not 5 (budget; disclosed)
• glm53.k6-6bpw-stream.brandonmusic-final25 non_sealed_lane: Produced by the 'streaming' lane, not the sealed-ep8 lane. On this panel the lane's offset against the sealed lane IS measured: -8.495843104593809e-06 nats on the mean (max 0.00028735280093581186 on any one window over 25 windows), and the tokenwise KL array is NOT the sealed one, so the run is not a reproduction of the sealed number.
• glm53.k6-6bpw-stream.brandonmusic-final25 note: Provenance of the fields the summary receipt does not carry. metric.direction and estimator.accumulation_dtype: SUPPLIED -- the k6-stream summary states neither, and both are recorded as the sealed lane's because the scorer is the same unmodified tools/k6_kld_report.py, invoked as --profile k6-stream. measurement_scope.contexts: READ from the verdict receipt's 25-entry per_window array, whose streaming means average to exactly the summary's measured_mean_kld. scored_positions: SUPPLIED as the panel's own 51,175 (25 x 2047), which the equal-weighted window average is consistent with. determinism.identical_across_runs: RECOMPUTED from run_means and distinct_tokenwise_kld_sha256; the receipt's bitwise_deterministic flag was checked against that, not copied. The verdict's sealed_mean_kld is bit-identical to the sealed K6 row in this file, which is what makes the delta a comparison of these two rows and not of two unrelated numbers. comparability.bias.floor_measurement_ref: SUPPLIED by --floor-measurement once the streaming-lane floor row below existed; build_row checked it was measured on this SAME lane before writing the reference (exit 7 otherwise).
• glm53.brandonmusic-4bpw.brandonmusic-final25 author_reported_only: Measured and published by brandonmusic on his own stack. We have not re-run it. It is nonetheless unusually well anchored: his receipt's token_panel_receipt_sha256 (0beec577...) and teacher_receipt_sha256 (2ae08117...) are byte-identical to ours, so the panel and the teacher are provably the same. Only the reader differs (1fb3be87... vs our 1ccce446...).
• glm53.brandonmusic-4bpw.brandonmusic-final25 note: On the single-window sub-panel the same artifact reads 0.022751 -- a 7% swing from 0.024555 over the full 25 windows.
• glm53.tr3-4bpw-stream.brandonmusic-final25 byte_identical_redistribution: The measured bytes are brandonmusic's, redistributed: all 120 shards have the same LFS oid as brandonmusic/GLM-5.3-Flash-tr3-4bpw @ 5ab363a8. The mirror was measured rather than the upstream because it pins a revision and the upstream record carries none. Credit for the quantization is brandonmusic's; credit for this number is ours.
• glm53.tr3-4bpw-stream.brandonmusic-final25 routed_experts_only_scope: scope glm53_routed_experts_only, non_routed_dtype_policy official_source_native, head_bits 16, read from the release's own config. Only the routed experts are quantized; all 1,618 non-routed tensors including lm_head are the OFFICIAL ones, verified name-set-equal to the official release's. The stock-exllamav3 rows on this same panel quantize attention, the dense MLPs, the shared experts, the vision tower and the head as well: at ~the same nominal bpw they are measuring a different amount of model.
• glm53.tr3-4bpw-stream.brandonmusic-final25 reduced_run_count: cold_run_deviation (verbatim from the receipt): 2 cold runs, not 5 (budget; disclosed)
• glm53.tr3-4bpw-stream.brandonmusic-final25 non_sealed_lane: Produced by the 'streaming' lane, not the sealed-ep8 lane. Unlike every other streaming row here, this artifact HAS a sealed-lane sibling to bridge against, because the bytes are provably identical to brandonmusic's: the same weights read 0.024554564249958208 there. The +0.000948863 nats between them is a LANE-PLUS-STACK offset, not a lane offset -- his run used his reader (1fb3be87...) and ours uses ours (1ccce446...) -- so it bounds the lane term rather than measuring it.
• glm53.tr3-4bpw-stream.brandonmusic-final25 note: First tr3-published artifact measurable by this suite: the streaming lane gained a reader (stream_score --source tr3, k6/tools/tr3_surface.py) in the same change. The routed decode is the campaign's own -- exl3hf_surface.decode_module over the frozen MCG LUT, proven bitwise identical to calling it directly -- so the codec path is the one the K6/K8 rows on this lane were measured through. The non-routed weights are the ARTIFACT's own, re-sharded VERBATIM by the materializer (1,618 tensors copied, 0 decoded, dtypes preserved) because they share shards with the 148,608 routed payload objects and transformers keys its checkpoint load off the shard files. No official-release weight is in the measured function. 907200 K4 expert matrices were decoded per cold run. Attributable error against this lane's own floor: 0.013997505 nats, versus 0.014020504 for turboderp's 4.05bpw -- the TR3 quant is the tighter of the two at ~the same nominal rate, on a strictly smaller quantized scope.
• glm53.turbo-4.05bpw-stream.brandonmusic-final25 unsealed_source: seal_disclosure (verbatim from the receipt): unsealed-source scoring: stock exllamav3 releases ship no upstream receipts, reconstruction closures or sealed reader ABI; the packed surface was decoded WITHOUT seal verification (consumed payload sha256s and the immutable repo revision are recorded instead)
• glm53.turbo-4.05bpw-stream.brandonmusic-final25 reduced_run_count: cold_run_deviation (verbatim from the receipt): 2 cold runs, not 5 (budget; disclosed)
• glm53.turbo-4.05bpw-stream.brandonmusic-final25 quantized_head: declared_head_bits 6 (verbatim from the receipt): this artifact's lm_head is itself quantized by the producer, unlike the TR3 artifacts on this panel which keep it native BF16. It is APPLIED natively from the artifact's own weights -- no shared or replayed head -- so estimator.head_policy is native_head; the quantization is artifact identity.
• glm53.turbo-4.05bpw-stream.brandonmusic-final25 non_sealed_lane: Produced by the 'streaming' lane, not the sealed-ep8 lane. The lane's offset against the sealed lane is NOT measured for this artifact: no sealed-lane row for it exists to bridge against.
• glm53.turbo-4.05bpw-stream.brandonmusic-final25 note: First artifact measured end to end by bin/measure-cloud. The receipt's family name carries no lane marker, so 'streaming' is SUPPLIED by --lane; direction, accumulation dtype, scored positions and context count are supplied too (this family is a scalar summary). determinism.identical_across_runs is RECOMPUTED from run_means and distinct_tokenwise_kld_sha256. The non-routed weights are the ARTIFACT's own, dequantized from its shards -- including its 6-bit head -- so no official-release weight is in the measured function; the materialization receipt is 3653c55f0dc729c3fccc6bbe5d8949b55e27517ade5d8c546fec79de03dd1c81. 907,200 K4 expert matrices were decoded per cold run. Top-1 agreement 0.9509916951636541, identical across both cold runs, read from the per-run kld-report.json (the scalar summary family did not carry it at the time this row was written; k6_kld_report now emits it).
• glm53.dione-q4.brandonmusic-final25 unsealed_source: The Dione checkpoint ships no upstream receipts or sealed reader ABI. The packed surface was decoded without seal verification; the immutable revision 99cccdf0... and the consumed payload sha256s were recorded instead (dione_shard_hash_verification: full).
• glm53.dione-q4.brandonmusic-final25 artifact_identity_incomplete: The release's own scope manifest was not parsed into this registry, so the artifact's per-class recipe is recorded as unknown.
• glm53.dione-q4.brandonmusic-final25 note: The receipt's cold_run_deviation field reads verbatim '5 cold runs, not 5 (budget; disclosed)' -- a self-contradictory template string. cold_run_count is 5 and run_means has 5 entries, so five runs is what happened; the string is a receipt-generator defect and is recorded here rather than copied into a disclosure. No clean17 sibling: receipt reports/dione-q4-packed-kld.json is scalar-only (no per_window block), so the calibration-clean scope cannot be recomputed without re-running the measurement.
• glm53.dione-3.0bpw-stream.brandonmusic-final25 unsealed_source: The Dione checkpoint ships no upstream receipts, reconstruction closures or sealed reader ABI, so the packed surface was decoded WITHOUT seal verification. What the release DOES publish is a per-shard sha256 manifest (EXL3_MANIFEST.json), and all 130 shard digests were recomputed on the measurement instance before anything was decoded (dione_shard_hash_verification: full); that, the immutable revision, the local config/index digests and the consumed-payload sha256 census are the provenance anchors.
• glm53.dione-3.0bpw-stream.brandonmusic-final25 reduced_run_count: cold_run_deviation (verbatim from the receipt): 2 cold runs, not 5 (budget; disclosed)
• glm53.dione-3.0bpw-stream.brandonmusic-final25 non_sealed_lane: Produced by the 'streaming' lane, not the sealed-ep8 lane. The lane's offset against the sealed lane is NOT measured for this artifact: no sealed-lane row for it exists to bridge against, and the offset its 4-bpw sibling would give is a property of the routing rather than a constant of the ladder.
• glm53.dione-3.0bpw-stream.brandonmusic-final25 note: Third rung of 0xSero's ladder measured here. His Q4 reads 0.027262784814670614 on the SEALED lane and this 3.0bpw reads 0.050501241465423556 on the STREAMING lane; the two are not directly comparable (different lane, different comparability key) and the registry refuses to net them. Within this lane the excess over the BF16-floor control (formerly: attributable error; P1-05) is 0.03899531884609326 nats. The producer's own RELEASE_STATUS.json marks this release quality: FAIL at their own threshold (their held-out forward KL 0.15251, top-1 0.87285 over 65,504 positions of THEIR panel) -- their number, their panel, their estimator, recorded on the artifact record rather than mixed into this one.
• glm53.turbo-2.05bpw-stream.brandonmusic-final25 unsealed_source: seal_disclosure (verbatim from the receipt): unsealed-source scoring: stock exllamav3 releases ship no upstream receipts, reconstruction closures or sealed reader ABI; the packed surface was decoded WITHOUT seal verification.
• glm53.turbo-2.05bpw-stream.brandonmusic-final25 reduced_run_count: cold_run_deviation (verbatim from the receipt): 2 cold runs, not 5 (budget; disclosed)
• glm53.turbo-2.05bpw-stream.brandonmusic-final25 quantized_head: declared_head_bits 5 -- lower than the 6 this producer's 4.05bpw and 3.05bpw branches declare. Applied natively from the artifact's own weights, so estimator.head_policy is native_head.
• glm53.turbo-2.05bpw-stream.brandonmusic-final25 non_sealed_lane: Produced by the 'streaming' lane, not the sealed-ep8 lane. The lane's offset against the sealed lane is NOT measured for this artifact.
• glm53.vcruz-k2-2bpw-stream.brandonmusic-final25 quality_gate_failed: The panel's gate is mean tokenwise KLD < 0.06 and this row reads 0.15520955491423008 -- 2.6x the threshold, and 3.07x the 3.0-bpw rung immediately above it. The gate is the artifact's verdict, not the measurement's: the run is bitwise deterministic, its two cold runs agree to the last bit, and the number is published exactly as it came out. What it says is that a 2-bit routed-expert quantization of this model diverges by 0.155 nats from its own BF16 source on this panel, at 87.27 % top-1 agreement.
• glm53.vcruz-k2-2bpw-stream.brandonmusic-final25 unsealed_source: The release ships no upstream receipts, no reconstruction closures, no sealed reader ABI -- and no per-shard digest list of its own: no SHA256SUMS, no EXL3_MANIFEST.json. What binds the bytes is the immutable 40-hex revision, the Hub's own per-file LFS content digests at that revision -- a 122-entry list, manifest digest 43a162282c06b19d098029afea4bedc77238026bca28fc514e50e33d827a9b66, captured from the models API BEFORE the rental and recomputed on the instance against the downloaded tree: 122/122 verified, 97,764,515,699 bytes, 0 absent, 0 safetensors on disk uncovered by the list -- plus the artifact's config sha256 163bd0888684f7eaf963ad67cdff3fbdca0749796c0aa5a6e7035816e503ecfc and index sha256 e9dd7cb2f6358843de334baa40ff537b4914721dbaa9c7dab42a386562afce19 recomputed locally and bound into the materialization receipt, and the consumed-payload sha256 census.
• glm53.vcruz-k2-2bpw-stream.brandonmusic-final25 reduced_run_count: cold_run_deviation (verbatim from the receipt): 2 cold runs, not 5 (budget; disclosed)
• glm53.vcruz-k2-2bpw-stream.brandonmusic-final25 non_sealed_lane: Produced by the 'streaming' lane, not the sealed-ep8 lane. The lane's offset against the sealed lane is NOT measured for this artifact: it has no sealed-lane row, and no sibling on either lane, to bridge against.
• glm53.vcruz-k2-2bpw-stream.brandonmusic-final25 note: The lowest rate measured on this panel, and the first row here to FAIL the 0.06 gate. 0.15520955491423008 nats at 87.27 % top-1, against 0.050501241465423556 at 93.00 % for the 3.0-bpw rung and 0.025503427634363770 at 95.31 % for 4 bpw: 3.07x the divergence of 3 bpw for 35 % fewer bytes (97.8 GB against 149.6 GB), and 6.09x the divergence of 4 bpw for 44 % fewer bytes. Against this lane's own BF16 floor the excess over control (formerly: quantization-attributable error; P1-05) is 0.143703632294899769 nats. Both cold runs produced identical run means and ONE tokenwise KL digest, so the path is bitwise deterministic; the divergence is the codec, not the harness. Every one of the 907,200 decoded expert matrices was K2 (routed_bits_decode_histogram {K2: 907200}), which is the decode side confirming the release's declared routed-experts-only scope. Per-domain the damage is uneven: axis2_legal 0.2509, axis1_general 0.1727, axis3_code_agentic 0.1272, axis4_reasoning_termination 0.0671 -- a 3.7x spread across domains that the single panel mean hides.

Group cmp--f0823827adb15376 -- 2 rows

Panel panel--glm53.brandonmusic.final25 -- brandonmusic GLM-5.3-Flash sealed qualification panel v1 -- 25 final windows 25 contexts x 2047 scored positions = 51,175 scored positions, score_from 0 sealed: yes (token digest 6bafe3283c54bc93...) -- contamination scan: NOT RUN Reference (teacher) reference--malaiwah.glm53-bf16-hf.brandonmusic-final25 -- native_bf16, artifact artifact--zai-org.glm-5.3-flash-bf16.a6c167b6 @a6c167b62691b2bac901344b65cb651a70f53e43 Metric mean_tokenwise_kld, direction reference_to_candidate, accumulation float64 Estimation surface stack_relation same_stack, head_policy native_head Comparability key cmp--f0823827adb15376 Like-for-like predicate comparable: false -- a RECORDED secondary dimension differs across members: replay_env. Equal keys make these rows candidates for comparison, not certified like-for-like; ranking across the differing dimension attributes a lane/pipeline/hardware/scope effect to quantization quality. Machine-readable form with per-dimension values: this key's comparability block in index.json.

What this table is. Every row here shares the comparability key above: the same tokens, the same teacher capture, the same metric and direction, the same estimator precision, the same stack relation and the same head policy. That makes them CANDIDATES for ranking -- the key is a necessary partition, not a certificate. Whether they are also like-for-like on the dimensions the key omits (lane, pipeline, scope coverage, hardware) is what the predicate line above answers.

Rank is not a verdict. The table is sorted by fidelity alone, and fidelity buys bits: a larger, higher-bitrate quant will usually sit above a smaller one, which is not news. Read the Size and Codec columns before reading the order, and compare like against like.

What it is NOT comparable to. The nearest neighbouring groups differ in:

• cmp--4a8630bdcadab97f (2 rows): reference_id reference--malaiwah.glm53-bf16-hf.brandonmusic-final25 -> reference--brandonmusic.glm53-bf16-fp32-logits.final25; stack_relation same_stack -> cross_stack
• cmp--202b717f3219c414 (11 rows): reference_id reference--malaiwah.glm53-bf16-hf.brandonmusic-final25 -> reference--brandonmusic.glm53-bf16-fp32-logits.final25; metric_name mean_tokenwise_kld -> mean_of_run_means_tokenwise_kld
• cmp--18990ab191ea7a67 (2 rows): panel_id panel--glm53.brandonmusic.final25 -> panel--glm53.brandonmusic.final-0000; reference_id reference--malaiwah.glm53-bf16-hf.brandonmusic-final25 -> reference--brandonmusic.glm53-bf16-fp32-logits.final-0000

Those numbers are in this file, under their own headings. Quoting one under the other heading is the mistake this layout exists to prevent: the key is a function of the panel, the teacher, the metric, the direction and the estimator, and the validator recomputes it from those fields rather than trusting the stamped value. What that catches is a row filed under a key its own fields do not produce. It does not catch a number attributed to the wrong panel in the first place -- no offline checker can. That is what the receipt digests on every row are for.

Also, and always: every table for a different model. A KL number is a divergence over one model's own vocabulary against that model's own teacher. It is not a quality score that can be carried between models.

Artifact	Codec	Size	mean_tokenwise_kld (nats)	CI95	Top-1	Runs	Attribution	Receipt
GLM-5.3-Flash BF16 @a6c167b6 (measurement floor)	bf16	--	0	--	100.00 %	2 runs, bitwise identical	measured by us (their artifact)	receipt
wrldsuksgo2mars GLM-5.3-Flash EXL3 K3 v1 (routed experts trellis K3 mcg, rest bf16)	exl3-mcg @3	136.7 GB	0.0505687	--	93.09 %	2 runs, bitwise identical	measured by us (their artifact)	receipt

The same artifact, measured elsewhere in this file. One of the artifacts below also carries a number in another table -- on a different panel, teacher or estimator -- and the widest of those spans 0%. None of the readings is wrong and none is interchangeable with another. Quoting one of them as the number for the artifact, without its table, is the misuse this registry exists to make obvious.

• GLM-5.3-Flash BF16 @a6c167b6 -- 4 values here, from 0 to 0.0127116 nats (0% apart). Other tables: cmp--202b717f3219c414, cmp--4a8630bdcadab97f, cmp--eee09298c558ab21.

Disclosures for the rows above (2)

• glm53-hf.exl3-k3-wrldsuksgo2mars.brandonmusic-final25 lossy_capture_codec: RECONSTRUCTED, NOT EXECUTED. The 36,288 routed-expert trellis payload groups of layers 3-44 are decoded to bf16 per module on the capture device (exl3-trellis-decode-to-bf16: exllamav3's unpack, tile permutation, two Hadamard GEMMs and su/sv scaling, mcg codebook read from each module's own marker, TF32 pinned off and recorded) BEFORE the loader; every non-routed tensor is carried as shipped. The decoder reproduces engines/tools/exl3hf_surface.py:decode_payload_hf bitwise on real payloads and in-house fp64 routes; it has NOT been proven bitwise against a running exllamav3 kernel, which is why this row is advisory.
• glm53-hf.exl3-k3-wrldsuksgo2mars.brandonmusic-final25 note: Per-window mean 0.050568748291117058, population sd 0.031431999700895066, min 0.013551408128013679 (final-0014, axis3_code_agentic), max 0.14424928435077825 (final-0004, axis1_general) over 25 windows; the token mean is the published value. NEW GROUP: scored against the same-lane root reference--malaiwah.glm53-bf16-hf.brandonmusic-final25, not against brandonmusic's teacher logits; do not read it beside the 13 older Flash rows on this panel.

Group cmp--4a8630bdcadab97f -- 2 rows

Panel panel--glm53.brandonmusic.final25 -- brandonmusic GLM-5.3-Flash sealed qualification panel v1 -- 25 final windows 25 contexts x 2047 scored positions = 51,175 scored positions, score_from 0 sealed: yes (token digest 6bafe3283c54bc93...) -- contamination scan: NOT RUN Reference (teacher) reference--brandonmusic.glm53-bf16-fp32-logits.final25 -- native_bf16, artifact artifact--zai-org.glm-5.3-flash-bf16.a6c167b6 @a6c167b62691b2bac901344b65cb651a70f53e43 Metric mean_tokenwise_kld, direction reference_to_candidate, accumulation float64 Estimation surface stack_relation cross_stack, head_policy native_head Comparability key cmp--4a8630bdcadab97f Like-for-like predicate comparable: unknown -- no recorded difference, but harness, lane, replay_backend, replay_env, stack are unrecorded for at least one member, so homogeneity cannot be certified. Machine-readable form with per-dimension values: this key's comparability block in index.json.

What this table is. Every row here shares the comparability key above: the same tokens, the same teacher capture, the same metric and direction, the same estimator precision, the same stack relation and the same head policy. That makes them CANDIDATES for ranking -- the key is a necessary partition, not a certificate. Whether they are also like-for-like on the dimensions the key omits (lane, pipeline, scope coverage, hardware) is what the predicate line above answers.

Rank is not a verdict. The table is sorted by fidelity alone, and fidelity buys bits: a larger, higher-bitrate quant will usually sit above a smaller one, which is not news. Read the Size and Codec columns before reading the order, and compare like against like.

What it is NOT comparable to. The nearest neighbouring groups differ in:

• cmp--f0823827adb15376 (2 rows): reference_id reference--brandonmusic.glm53-bf16-fp32-logits.final25 -> reference--malaiwah.glm53-bf16-hf.brandonmusic-final25; stack_relation cross_stack -> same_stack
• cmp--eee09298c558ab21 (2 rows): panel_id panel--glm53.brandonmusic.final25 -> panel--glm53.brandonmusic.final25-clean17; reference_id reference--brandonmusic.glm53-bf16-fp32-logits.final25 -> reference--brandonmusic.glm53-bf16-fp32-logits.final25-clean17
• cmp--202b717f3219c414 (11 rows): metric_name mean_tokenwise_kld -> mean_of_run_means_tokenwise_kld; stack_relation cross_stack -> same_stack

Those numbers are in this file, under their own headings. Quoting one under the other heading is the mistake this layout exists to prevent: the key is a function of the panel, the teacher, the metric, the direction and the estimator, and the validator recomputes it from those fields rather than trusting the stamped value. What that catches is a row filed under a key its own fields do not produce. It does not catch a number attributed to the wrong panel in the first place -- no offline checker can. That is what the receipt digests on every row are for.

Also, and always: every table for a different model. A KL number is a divergence over one model's own vocabulary against that model's own teacher. It is not a quality score that can be carried between models.

Artifact	Codec	Size	mean_tokenwise_kld (nats)	CI95	Top-1	Runs	Attribution	Receipt
GLM-5.3-Flash BF16 @a6c167b6 (measurement floor)	bf16	--	0.0127116	[0.0103287, 0.0153918]	96.65 %	1 run, unevidenced	measured by us (their artifact)	receipt
GLM-5.3-Flash official FP8	fp8_e4m3 @8	328.4 GB	0.0206153	[0.0164689, 0.0256239]	95.63 %	1 run, unevidenced	measured by us (their artifact)	receipt

Bias on GLM-5.3-Flash BF16 @a6c167b6 -- cross_stack_capture_replay, direction upward. THIS ROW IS THE FLOOR. It replays the reference's own BF16 weights through our vLLM stack and scores them against brandonmusic's stored fp32 teacher logits. 0.012712 nats is therefore what two stacks disagree by on identical unquantized weights -- not a quantization result. No floor is named because none exists below it.

Bias on GLM-5.3-Flash official FP8 -- cross_stack_capture_replay, direction unknown. Teacher captured on brandonmusic's transformers/eager stack, candidate replayed on our vLLM stack. The matched unquantized control on this exact panel is 0.012712. The raw difference is 0.007904, a signed descriptive contrast, not a causal estimate or an upper bound on native-serving divergence. KL is not additive.

The same artifact, measured elsewhere in this file. 2 of the artifacts below also carry a number in another table -- on a different panel, teacher or estimator -- and the widest of those spans 51%. None of the readings is wrong and none is interchangeable with another. Quoting one of them as the number for the artifact, without its table, is the misuse this registry exists to make obvious.

• GLM-5.3-Flash BF16 @a6c167b6 -- 4 values here, from 0 to 0.0127116 nats (0% apart). Other tables: cmp--202b717f3219c414, cmp--eee09298c558ab21, cmp--f0823827adb15376.
• GLM-5.3-Flash official FP8 -- 4 values here, from 0.0186653 to 0.0281039 nats (51% apart). Other tables: cmp--9b009314102d9e8b, cmp--e6cdd07242bdde05, cmp--eee09298c558ab21.

Disclosures for the rows above (5)

• glm53.bf16-replay-floor.brandonmusic-final25 cross_stack_capture: Teacher captured on transformers/eager (B200 x4); candidate replayed on our vLLM stack. The offset audit confirms position alignment: top-1 agreement is 0.9665 at offset 0 and 0.0159 / 0.0162 at offsets -1 / +1.
• glm53.bf16-replay-floor.brandonmusic-final25 single_run: One pass; determinism not established.
• glm53.bf16-replay-floor.brandonmusic-final25 note: CONTROL ROW / MEASUREMENT FLOOR. Every cross-stack row on this panel contains this term.
• glm53.official-fp8.brandonmusic-final25.crossstack cross_stack_capture: This row cannot be ranked against the K6 / Dione / 4bpw rows on the same panel: those are same-stack sealed-capture numbers and this is a cross-stack replay. Their comparability keys differ, and the registry's tables are grouped by that key.
• glm53.official-fp8.brandonmusic-final25.crossstack single_run: One pass; determinism not established.

Panel: brandonmusic panel v1, calibration-clean subset -- 17 of 25 final windows

Derived from panel--glm53.brandonmusic.final25 by shard_subset: The 17 of 25 sealed windows whose 13-gram overlap with the calibration-role windows is at or below 0.05. Dropped: final-0003, final-0007, final-0011, final-0015, final-0019, final-0021, final-0022, final-0023. Six of the eight are axis4_reasoning_termination at 37-39% overlap, which removes that domain entirely; the other two (final-0021 at 7.1%, final-0022 at 5.8%) are legal and code-agentic, so this is NOT a whole-domain drop and a 19-window axis4-only exclusion is a DIFFERENT scope. The highest overlap among the retained windows is 4.75% (final-0014), so the 5% threshold separates cleanly but not by much -- it is inherited from brandonmusic and is an open joint decision, not a derived constant.

Panel disclosure -- subset_of_panel: 17 of the parent panel's 25 windows, 34799 of 51,175 scored positions. Rows on this panel must never be tabled beside rows on the parent panel: excluding the contaminated windows moves different contributors' numbers in OPPOSITE directions (every malaiwah row falls 12.6-16.2%, brandonmusic's own 4bpw row rises 1.6%).

Panel disclosure -- calibration_panel_overlap: This panel exists because the parent's contamination guard was role separation only. The n-gram scan that produced it found one whole domain sharing 37-39% of its 13-grams with calibration-role windows despite clean document-level separation -- the finding is brandonmusic's and the reproduction is ours.

This panel carries 2 separate comparability groups. They are different measurements of different things and are never merged.

Group cmp--2b9c401d13806d7e -- 4 rows

Panel panel--glm53.brandonmusic.final25-clean17 -- brandonmusic panel v1, calibration-clean subset -- 17 of 25 final windows 17 contexts x 2047 scored positions = 34,799 scored positions, score_from 0 sealed: yes (token digest ecfce5997ab9106c...) -- contamination scan: yes, 8 hits Reference (teacher) reference--brandonmusic.glm53-bf16-fp32-logits.final25-clean17 -- native_bf16, artifact artifact--zai-org.glm-5.3-flash-bf16.a6c167b6 @a6c167b62691b2bac901344b65cb651a70f53e43 Metric mean_of_run_means_tokenwise_kld, direction reference_to_candidate, accumulation float64 Estimation surface stack_relation same_stack, head_policy native_head Comparability key cmp--2b9c401d13806d7e Like-for-like predicate comparable: false -- a RECORDED secondary dimension differs across members: lane, pipeline. Equal keys make these rows candidates for comparison, not certified like-for-like; ranking across the differing dimension attributes a lane/pipeline/hardware/scope effect to quantization quality. Machine-readable form with per-dimension values: this key's comparability block in index.json.

What this table is. Every row here shares the comparability key above: the same tokens, the same teacher capture, the same metric and direction, the same estimator precision, the same stack relation and the same head policy. That makes them CANDIDATES for ranking -- the key is a necessary partition, not a certificate. Whether they are also like-for-like on the dimensions the key omits (lane, pipeline, scope coverage, hardware) is what the predicate line above answers.

Rank is not a verdict. The table is sorted by fidelity alone, and fidelity buys bits: a larger, higher-bitrate quant will usually sit above a smaller one, which is not news. Read the Size and Codec columns before reading the order, and compare like against like.

What it is NOT comparable to. The nearest neighbouring groups differ in:

• cmp--eee09298c558ab21 (2 rows): metric_name mean_of_run_means_tokenwise_kld -> mean_tokenwise_kld; stack_relation same_stack -> cross_stack
• cmp--202b717f3219c414 (11 rows): panel_id panel--glm53.brandonmusic.final25-clean17 -> panel--glm53.brandonmusic.final25; reference_id reference--brandonmusic.glm53-bf16-fp32-logits.final25-clean17 -> reference--brandonmusic.glm53-bf16-fp32-logits.final25

Those numbers are in this file, under their own headings. Quoting one under the other heading is the mistake this layout exists to prevent: the key is a function of the panel, the teacher, the metric, the direction and the estimator, and the validator recomputes it from those fields rather than trusting the stamped value. What that catches is a row filed under a key its own fields do not produce. It does not catch a number attributed to the wrong panel in the first place -- no offline checker can. That is what the receipt digests on every row are for.

Also, and always: every table for a different model. A KL number is a divergence over one model's own vocabulary against that model's own teacher. It is not a quality score that can be carried between models.

2 of this group's 4 rows came off a different measurement lane (streaming) and are tabled on their own below, not mixed into the ordering here. The key does not carry the lane; this file does.

Artifact	Codec	Size	mean_of_run_means_tokenwise_kld (nats)	CI95	Top-1	Runs	Attribution	Receipt
malaiwah GLM-5.3-Flash TR3 6bpw (K6)	exl3-mcg @6	253.5 GB	0.0116773	[0.00885591, 0.0147851]	--	5 runs, bitwise identical	measured by us	receipt
brandonmusic GLM-5.3-Flash tr3 4bpw	exl3-mcg @4	175.6 GB	0.0249488	[0.0181407, 0.0416622]	--	5 runs, bitwise identical	reported by brandonmusic	receipt

Lane streaming -- 2 of this group's 4 rows

A different lane. Same key, and that is exactly the problem this table solves. The comparability key is a function of the panel, the teacher, the metric, the direction, the estimator precision, the stack relation and the head policy -- and these rows match the table above on all seven. What they do not share is the machine and the code path that produced the candidate logits, and lanes are not interchangeable. Sorting them into one list would read as a ranking; where the same artifact appears in both, it is one set of weights measured twice, not two quants.

What the lane is (from pipeline--malaiwah.glm53-stream-packed-kld): 1 device, expert-parallel width 8 emulated in one process, routed-expert combine order fp32.

Bridge to the sealed-ep8 lane, measured on this panel: signed delta -8.4958e-06 nats on the mean against measurement--glm53.k6-6bpw.brandonmusic-final25, worst single window 0.00028735, over 25 windows. Tokenwise KL array matches the sealed run: no. The runner's own verdict on whether this may be published as a reproduction of the sealed number: no (verdict LARGER_DELTA_SEE_DISCLOSURE).

That bridge is one artifact's, on one panel. It is not a constant and it is not subtractable: a row in this table whose artifact has no sealed-lane row has no measured offset at all, and says so in its own bias line.

Artifact	Codec	Size	mean_of_run_means_tokenwise_kld (nats)	CI95	Top-1	Runs	Attribution	Receipt
malaiwah GLM-5.3-Flash TR3 8bpw (K8)	exl3-mcg @8	331.4 GB	0.0108294	[0.0080632, 0.0138378]	--	2 runs, bitwise identical	measured by us	receipt
malaiwah GLM-5.3-Flash TR3 6bpw (K6)	exl3-mcg @6	253.5 GB	0.011676	[0.00886271, 0.0147792]	--	2 runs, bitwise identical	measured by us	receipt

Bias on malaiwah GLM-5.3-Flash TR3 8bpw (K8) -- other, direction unknown. NO FLOOR ON THIS SCOPE: the same-lane floor (measurement--glm53.bf16-stream-floor.brandonmusic-final25) has a scalar-only receipt with no per-window array, so it cannot be recomputed on the calibration-clean window set. This row has no floor reference; panel25 floor values do not apply to clean17.

Bias on malaiwah GLM-5.3-Flash TR3 6bpw (K6) -- other, direction downward. NO FLOOR ON THIS SCOPE: the same-lane floor (measurement--glm53.bf16-stream-floor.brandonmusic-final25) has a scalar-only receipt with no per-window array, so it cannot be recomputed on the calibration-clean window set. This row has no floor reference; panel25 floor values do not apply to clean17.

The same artifact, measured elsewhere in this file. 3 of the artifacts below also carry a number in another table -- on a different panel, teacher or estimator -- and the widest of those spans 18%. None of the readings is wrong and none is interchangeable with another. Quoting one of them as the number for the artifact, without its table, is the misuse this registry exists to make obvious.

the 3 artifacts and their ranges

• brandonmusic GLM-5.3-Flash tr3 4bpw -- 3 values here, from 0.0227508 to 0.0249488 nats (10% apart). Other tables: cmp--18990ab191ea7a67, cmp--202b717f3219c414.
• malaiwah GLM-5.3-Flash TR3 6bpw (K6) -- 6 values here, from 0.011676 to 0.0137234 nats (18% apart). Other tables: cmp--202b717f3219c414.
• malaiwah GLM-5.3-Flash TR3 8bpw (K8) -- 2 values here, from 0.0108294 to 0.0123842 nats (14% apart). Other tables: cmp--202b717f3219c414.

Disclosures for the rows above (13)

• glm53.k8-8bpw-stream.brandonmusic-final25.clean17 reduced_run_count: cold_run_deviation (verbatim from the receipt): 2 cold runs, not 5 (budget; disclosed)
• glm53.k8-8bpw-stream.brandonmusic-final25.clean17 non_sealed_lane: Produced by the 'streaming' lane, not the sealed-ep8 lane. The lane's offset against the sealed lane is NOT measured for this artifact: no sealed-lane row for it exists to bridge against.
• glm53.k8-8bpw-stream.brandonmusic-final25.clean17 subset_of_panel: 17 of the panel's 25 sealed windows (34,799 of 51,175 scored positions). The excluded 8 are the windows the calibration-overlap scan flags; see measurement_scope.calibration_overlap_scan.
• glm53.k8-8bpw-stream.brandonmusic-final25.clean17 note: Calibration-clean scope recompute of measurement--glm53.k8-8bpw-stream.brandonmusic-final25. panel25 0.012384191023437 -> clean17 0.010829419869883 (-0.001554771153554, -12.55%). Not a correction and not a supersession: the two scopes answer different questions and move different contributors' rows in opposite directions. Never compare a clean17 value against a panel25 value.
• glm53.k6-6bpw-stream.brandonmusic-final25.clean17 reduced_run_count: cold_run_deviation (verbatim from the receipt): 2 cold runs, not 5 (budget; disclosed)
• glm53.k6-6bpw-stream.brandonmusic-final25.clean17 non_sealed_lane: Produced by the 'streaming' lane, not the sealed-ep8 lane. On this panel the lane's offset against the sealed lane IS measured: -8.495843104593809e-06 nats on the mean (max 0.00028735280093581186 on any one window over 25 windows), and the tokenwise KL array is NOT the sealed one, so the run is not a reproduction of the sealed number.
• glm53.k6-6bpw-stream.brandonmusic-final25.clean17 subset_of_panel: 17 of the panel's 25 sealed windows (34,799 of 51,175 scored positions). The excluded 8 are the windows the calibration-overlap scan flags; see measurement_scope.calibration_overlap_scan.
• glm53.k6-6bpw-stream.brandonmusic-final25.clean17 note: Calibration-clean scope recompute of measurement--glm53.k6-6bpw-stream.brandonmusic-final25. panel25 0.013714888822597 -> clean17 0.011675992693735 (-0.002038896128862, -14.87%). Not a correction and not a supersession: the two scopes answer different questions and move different contributors' rows in opposite directions. Never compare a clean17 value against a panel25 value.
• glm53.k6-6bpw.brandonmusic-final25.clean17 subset_of_panel: 17 of the panel's 25 sealed windows (34,799 of 51,175 scored positions). The excluded 8 are the windows the calibration-overlap scan flags; see measurement_scope.calibration_overlap_scan.
• glm53.k6-6bpw.brandonmusic-final25.clean17 note: Calibration-clean scope recompute of measurement--glm53.k6-6bpw.brandonmusic-final25. panel25 0.013723384665701 -> clean17 0.011677286368695 (-0.002046098297006, -14.91%). Not a correction and not a supersession: the two scopes answer different questions and move different contributors' rows in opposite directions. Never compare a clean17 value against a panel25 value.
• glm53.brandonmusic-4bpw.brandonmusic-final25.clean17 author_reported_only: Measured and published by brandonmusic on his own stack. We have not re-run it. It is nonetheless unusually well anchored: his receipt's token_panel_receipt_sha256 (0beec577...) and teacher_receipt_sha256 (2ae08117...) are byte-identical to ours, so the panel and the teacher are provably the same. Only the reader differs (1fb3be87... vs our 1ccce446...).
• glm53.brandonmusic-4bpw.brandonmusic-final25.clean17 subset_of_panel: 17 of the panel's 25 sealed windows (34,799 of 51,175 scored positions). The excluded 8 are the windows the calibration-overlap scan flags; see measurement_scope.calibration_overlap_scan.
• glm53.brandonmusic-4bpw.brandonmusic-final25.clean17 note: Calibration-clean scope recompute of measurement--glm53.brandonmusic-4bpw.brandonmusic-final25. panel25 0.024554564249958 -> clean17 0.024948837055615 (+0.000394272805657, +1.61%). Not a correction and not a supersession: the two scopes answer different questions and move different contributors' rows in opposite directions. Never compare a clean17 value against a panel25 value.

Group cmp--eee09298c558ab21 -- 2 rows

Panel panel--glm53.brandonmusic.final25-clean17 -- brandonmusic panel v1, calibration-clean subset -- 17 of 25 final windows 17 contexts x 2047 scored positions = 34,799 scored positions, score_from 0 sealed: yes (token digest ecfce5997ab9106c...) -- contamination scan: yes, 8 hits Reference (teacher) reference--brandonmusic.glm53-bf16-fp32-logits.final25-clean17 -- native_bf16, artifact artifact--zai-org.glm-5.3-flash-bf16.a6c167b6 @a6c167b62691b2bac901344b65cb651a70f53e43 Metric mean_tokenwise_kld, direction reference_to_candidate, accumulation float64 Estimation surface stack_relation cross_stack, head_policy native_head Comparability key cmp--eee09298c558ab21 Like-for-like predicate comparable: unknown -- no recorded difference, but lane, replay_backend, replay_env, stack are unrecorded for at least one member, so homogeneity cannot be certified. Machine-readable form with per-dimension values: this key's comparability block in index.json.

What this table is. Every row here shares the comparability key above: the same tokens, the same teacher capture, the same metric and direction, the same estimator precision, the same stack relation and the same head policy. That makes them CANDIDATES for ranking -- the key is a necessary partition, not a certificate. Whether they are also like-for-like on the dimensions the key omits (lane, pipeline, scope coverage, hardware) is what the predicate line above answers.

Rank is not a verdict. The table is sorted by fidelity alone, and fidelity buys bits: a larger, higher-bitrate quant will usually sit above a smaller one, which is not news. Read the Size and Codec columns before reading the order, and compare like against like.

What it is NOT comparable to. The nearest neighbouring groups differ in:

• cmp--4a8630bdcadab97f (2 rows): panel_id panel--glm53.brandonmusic.final25-clean17 -> panel--glm53.brandonmusic.final25; reference_id reference--brandonmusic.glm53-bf16-fp32-logits.final25-clean17 -> reference--brandonmusic.glm53-bf16-fp32-logits.final25
• cmp--2b9c401d13806d7e (4 rows): metric_name mean_tokenwise_kld -> mean_of_run_means_tokenwise_kld; stack_relation cross_stack -> same_stack

Those numbers are in this file, under their own headings. Quoting one under the other heading is the mistake this layout exists to prevent: the key is a function of the panel, the teacher, the metric, the direction and the estimator, and the validator recomputes it from those fields rather than trusting the stamped value. What that catches is a row filed under a key its own fields do not produce. It does not catch a number attributed to the wrong panel in the first place -- no offline checker can. That is what the receipt digests on every row are for.

Also, and always: every table for a different model. A KL number is a divergence over one model's own vocabulary against that model's own teacher. It is not a quality score that can be carried between models.

Artifact	Codec	Size	mean_tokenwise_kld (nats)	CI95	Top-1	Runs	Attribution	Receipt
GLM-5.3-Flash BF16 @a6c167b6 (measurement floor)	bf16	--	0.0106476	[0.00812433, 0.0134906]	--	1 run, unevidenced	measured by us (their artifact)	receipt
GLM-5.3-Flash official FP8	fp8_e4m3 @8	328.4 GB	0.0186653	[0.0141924, 0.0247936]	--	1 run, unevidenced	measured by us (their artifact)	receipt

Bias on GLM-5.3-Flash BF16 @a6c167b6 -- cross_stack_capture_replay, direction upward. THIS ROW IS THE FLOOR. It replays the reference's own BF16 weights through our vLLM stack and scores them against brandonmusic's stored fp32 teacher logits. 0.012712 nats is therefore what two stacks disagree by on identical unquantized weights -- not a quantization result. No floor is named because none exists below it.

Bias on GLM-5.3-Flash official FP8 -- cross_stack_capture_replay, direction unknown. Clean17 descriptive cross-stack context only. The referenced floor is recomputed on the same 17 windows; the panel25 floor value does not apply here. A shared lane alone does not establish additive bias or authorize causal subtraction.

The same artifact, measured elsewhere in this file. 2 of the artifacts below also carry a number in another table -- on a different panel, teacher or estimator -- and the widest of those spans 51%. None of the readings is wrong and none is interchangeable with another. Quoting one of them as the number for the artifact, without its table, is the misuse this registry exists to make obvious.

• GLM-5.3-Flash BF16 @a6c167b6 -- 4 values here, from 0 to 0.0127116 nats (0% apart). Other tables: cmp--202b717f3219c414, cmp--4a8630bdcadab97f, cmp--f0823827adb15376.
• GLM-5.3-Flash official FP8 -- 4 values here, from 0.0186653 to 0.0281039 nats (51% apart). Other tables: cmp--4a8630bdcadab97f, cmp--9b009314102d9e8b, cmp--e6cdd07242bdde05.

Disclosures for the rows above (8)

• glm53.bf16-replay-floor.brandonmusic-final25.clean17 cross_stack_capture: Teacher captured on transformers/eager (B200 x4); candidate replayed on our vLLM stack. The offset audit confirms position alignment: top-1 agreement is 0.9665 at offset 0 and 0.0159 / 0.0162 at offsets -1 / +1.
• glm53.bf16-replay-floor.brandonmusic-final25.clean17 single_run: One pass; determinism not established.
• glm53.bf16-replay-floor.brandonmusic-final25.clean17 subset_of_panel: 17 of the panel's 25 sealed windows (34,799 of 51,175 scored positions). The excluded 8 are the windows the calibration-overlap scan flags; see measurement_scope.calibration_overlap_scan.
• glm53.bf16-replay-floor.brandonmusic-final25.clean17 note: Calibration-clean scope recompute of measurement--glm53.bf16-replay-floor.brandonmusic-final25. panel25 0.012711599817251 -> clean17 0.010647639361035 (-0.002063960456216, -16.24%). Not a correction and not a supersession: the two scopes answer different questions and move different contributors' rows in opposite directions. Never compare a clean17 value against a panel25 value.
• glm53.official-fp8.brandonmusic-final25.crossstack.clean17 cross_stack_capture: This row cannot be ranked against the K6 / Dione / 4bpw rows on the same panel: those are same-stack sealed-capture numbers and this is a cross-stack replay. Their comparability keys differ, and the registry's tables are grouped by that key.
• glm53.official-fp8.brandonmusic-final25.crossstack.clean17 single_run: One pass; determinism not established.
• glm53.official-fp8.brandonmusic-final25.crossstack.clean17 subset_of_panel: 17 of the panel's 25 sealed windows (34,799 of 51,175 scored positions). The excluded 8 are the windows the calibration-overlap scan flags; see measurement_scope.calibration_overlap_scan.
• glm53.official-fp8.brandonmusic-final25.crossstack.clean17 note: Calibration-clean scope recompute of measurement--glm53.official-fp8.brandonmusic-final25.crossstack. panel25 0.020615254540418 -> clean17 0.018665326569455 (-0.001949927970963, -9.46%). Not a correction and not a supersession: the two scopes answer different questions and move different contributors' rows in opposite directions. Never compare a clean17 value against a panel25 value.

Panel: brandonmusic panel v1, single window final-0000

Derived from panel--glm53.brandonmusic.final25 by shard_subset: window final-0000 alone, 1/25 of the parent panel. 2,047 scored positions instead of 51,175. brandonmusic's runtime receipts score this window only. The same artifact reads 0.022751 here and 0.024555 over the full 25 windows, a 7% swing -- which is why this is a separate panel record.

Panel disclosure -- weak_contamination_guard: This panel's only contamination guard is ROLE SEPARATION: the 25 'final' windows are drawn from the same packed corpus as the 384 fit / 128 conditional-fit / 64 selection / 64 confirmation windows and are declared qualification-only. No lexical or n-gram scan is published, and the underlying document provenance is published only as a digest. This is materially weaker than the malaiwah v5 suites, which run a 12-word shingle whole-document pre-exclusion and report 0 hits. Do not describe the two guards as equivalent. It applies equally to every row on this panel, so it does not disturb comparisons WITHIN the panel.

Panel disclosure -- subset_of_panel: A single 2,047-position window. Numbers on this panel have far wider sampling error than the 25-window panel and must never be tabled beside it.

This panel carries 2 separate comparability groups. They are different measurements of different things and are never merged.

Group cmp--b55c2d693d127f20 -- 6 rows

Panel panel--glm53.brandonmusic.final-0000 -- brandonmusic panel v1, single window final-0000 1 contexts x 2047 scored positions = 2,047 scored positions, score_from 0 sealed: yes (token digest 338027e62f41540f...) -- contamination scan: NOT RUN Reference (teacher) reference--brandonmusic.glm53-bf16-fp32-logits.final-0000 -- native_bf16, artifact artifact--zai-org.glm-5.3-flash-bf16.a6c167b6 @a6c167b62691b2bac901344b65cb651a70f53e43 Metric mean_of_run_means_tokenwise_kld, direction reference_to_candidate, accumulation unknown Estimation surface stack_relation same_stack, head_policy native_head Comparability key cmp--b55c2d693d127f20 Like-for-like predicate comparable: false -- a RECORDED secondary dimension differs across members: pipeline. Equal keys make these rows candidates for comparison, not certified like-for-like; ranking across the differing dimension attributes a lane/pipeline/hardware/scope effect to quantization quality. Machine-readable form with per-dimension values: this key's comparability block in index.json.

What this table is. Every row here shares the comparability key above: the same tokens, the same teacher capture, the same metric and direction, the same estimator precision, the same stack relation and the same head policy. That makes them CANDIDATES for ranking -- the key is a necessary partition, not a certificate. Whether they are also like-for-like on the dimensions the key omits (lane, pipeline, scope coverage, hardware) is what the predicate line above answers.

Rank is not a verdict. The table is sorted by fidelity alone, and fidelity buys bits: a larger, higher-bitrate quant will usually sit above a smaller one, which is not news. Read the Size and Codec columns before reading the order, and compare like against like.

What it is NOT comparable to. The nearest neighbouring groups differ in:

• cmp--18990ab191ea7a67 (2 rows): metric_name mean_of_run_means_tokenwise_kld -> mean_tokenwise_kld; accumulation_dtype unknown -> float64

Those numbers are in this file, under their own headings. Quoting one under the other heading is the mistake this layout exists to prevent: the key is a function of the panel, the teacher, the metric, the direction and the estimator, and the validator recomputes it from those fields rather than trusting the stamped value. What that catches is a row filed under a key its own fields do not produce. It does not catch a number attributed to the wrong panel in the first place -- no offline checker can. That is what the receipt digests on every row are for.

Also, and always: every table for a different model. A KL number is a divergence over one model's own vocabulary against that model's own teacher. It is not a quality score that can be carried between models.

Artifact	Codec	Size	mean_of_run_means_tokenwise_kld (nats)	CI95	Top-1	Runs	Attribution	Receipt
GLM-5.3-Flash official FP8 weights served with FP8 MLA KV	fp8_e4m3 @8	--	0.0245817	--	93.63 %	5 runs, sd 0.00016	reported by brandonmusic	receipt
GLM-5.3-Flash official FP8 weights served with FP8 MLA KV	fp8_e4m3 @8	--	0.0246106	--	93.73 %	5 runs, sd 0.000257	reported by brandonmusic	receipt
GLM-5.3-Flash official FP8 weights served with FP8 MLA KV	fp8_e4m3 @8	--	0.0246286	--	93.80 %	5 runs, sd 0.000326	reported by brandonmusic	receipt
brandonmusic GLM-5.3-Flash NVFP4 runtime build	nvfp4 @4	--	0.0547574	--	91.50 %	5 runs, bitwise identical	reported by brandonmusic	receipt
brandonmusic GLM-5.3-Flash NVFP4 runtime build	nvfp4 @4	--	0.0547574	--	91.50 %	5 runs, bitwise identical	reported by brandonmusic	receipt
brandonmusic GLM-5.3-Flash NVFP4 runtime build	nvfp4 @4	--	0.0605349	--	91.55 %	5 runs, bitwise identical	reported by brandonmusic	receipt

The same artifact, measured elsewhere in this file. One of the artifacts below also carries a number in another table -- on a different panel, teacher or estimator -- and the widest of those spans 25%. None of the readings is wrong and none is interchangeable with another. Quoting one of them as the number for the artifact, without its table, is the misuse this registry exists to make obvious.

• brandonmusic GLM-5.3-Flash NVFP4 runtime build -- 6 values here, from 0.0547574 to 0.0682296 nats (25% apart). Other tables: cmp--18990ab191ea7a67.

Disclosures for the rows above (13)

• glm53.official-fp8.v71.brandonmusic-final-0000 author_reported_only: Measured and published by brandonmusic on his own runtime image. Regime as published: FP8 MLA NoPE, route128 SMEM, TP2/EP2, DCP2 B12X A2A eager no-MTP. We have not re-run it.
• glm53.official-fp8.v71.brandonmusic-final-0000 estimator_unknown: This receipt family (glm53-r19-runtime-kld-repeated.v1) publishes no compute_dtype, so the accumulation precision of brandonmusic's scorer is not established for these rows and is recorded as unknown. All six rows in this group share that condition, so they remain mutually comparable; a row whose receipt attests float64 would not join them. His other two GLM-5.3-Flash receipts do declare float64, which makes it likely but not evidenced here.
• glm53.official-fp8.v75.brandonmusic-final-0000 author_reported_only: Measured and published by brandonmusic on his own runtime image. Regime as published: v75 release image, FP8 MLA NoPE, route128 SMEM/register, TP2/EP2, DCP2 direct symmetric-memory A2A. We have not re-run it.
• glm53.official-fp8.v75.brandonmusic-final-0000 estimator_unknown: This receipt family (glm53-r19-runtime-kld-repeated.v1) publishes no compute_dtype, so the accumulation precision of brandonmusic's scorer is not established for these rows and is recorded as unknown. All six rows in this group share that condition, so they remain mutually comparable; a row whose receipt attests float64 would not join them. His other two GLM-5.3-Flash receipts do declare float64, which makes it likely but not evidenced here.
• glm53.official-fp8.v44.brandonmusic-final-0000 author_reported_only: Measured and published by brandonmusic on his own runtime image. Regime as published: v43 TP2 DCP1 eager no-MTP FP8 MLA KV, GPUs 2,3. We have not re-run it.
• glm53.official-fp8.v44.brandonmusic-final-0000 estimator_unknown: This receipt family (glm53-r19-runtime-kld-repeated.v1) publishes no compute_dtype, so the accumulation precision of brandonmusic's scorer is not established for these rows and is recorded as unknown. All six rows in this group share that condition, so they remain mutually comparable; a row whose receipt attests float64 would not join them. His other two GLM-5.3-Flash receipts do declare float64, which makes it likely but not evidenced here.
• glm53.nvfp4.v71.brandonmusic-final-0000 author_reported_only: Measured and published by brandonmusic on his own runtime image. Regime as published: NVFP4 MLA NoPE, power-of-two ceil amax scale v2, route128 SMEM, TP2/EP2, DCP2 B12X A2A eager no-MTP. We have not re-run it.
• glm53.nvfp4.v71.brandonmusic-final-0000 estimator_unknown: This receipt family (glm53-r19-runtime-kld-repeated.v1) publishes no compute_dtype, so the accumulation precision of brandonmusic's scorer is not established for these rows and is recorded as unknown. All six rows in this group share that condition, so they remain mutually comparable; a row whose receipt attests float64 would not join them. His other two GLM-5.3-Flash receipts do declare float64, which makes it likely but not evidenced here.
• glm53.nvfp4.v75.brandonmusic-final-0000 author_reported_only: Measured and published by brandonmusic on his own runtime image. Regime as published: v75 release image, NVFP4 MLA NoPE calibrated power-of-two 46-layer scales. We have not re-run it.
• glm53.nvfp4.v75.brandonmusic-final-0000 estimator_unknown: This receipt family (glm53-r19-runtime-kld-repeated.v1) publishes no compute_dtype, so the accumulation precision of brandonmusic's scorer is not established for these rows and is recorded as unknown. All six rows in this group share that condition, so they remain mutually comparable; a row whose receipt attests float64 would not join them. His other two GLM-5.3-Flash receipts do declare float64, which makes it likely but not evidenced here.
• glm53.nvfp4.v44.brandonmusic-final-0000 author_reported_only: Measured and published by brandonmusic on his own runtime image. Regime as published: v44 TP2 DCP1 eager no-MTP NVFP4 MLA KV, GPUs 2,3. We have not re-run it.
• glm53.nvfp4.v44.brandonmusic-final-0000 estimator_unknown: This receipt family (glm53-r19-runtime-kld-repeated.v1) publishes no compute_dtype, so the accumulation precision of brandonmusic's scorer is not established for these rows and is recorded as unknown. All six rows in this group share that condition, so they remain mutually comparable; a row whose receipt attests float64 would not join them. His other two GLM-5.3-Flash receipts do declare float64, which makes it likely but not evidenced here.
• glm53.nvfp4.v44.brandonmusic-final-0000 quality_gate_failed: The author's own gate (mean tokenwise KLD < 0.06) did NOT pass. Recorded because a failing gate is a fact about the artifact, not a reason to hide the row.

Group cmp--18990ab191ea7a67 -- 2 rows

Panel panel--glm53.brandonmusic.final-0000 -- brandonmusic panel v1, single window final-0000 1 contexts x 2047 scored positions = 2,047 scored positions, score_from 0 sealed: yes (token digest 338027e62f41540f...) -- contamination scan: NOT RUN Reference (teacher) reference--brandonmusic.glm53-bf16-fp32-logits.final-0000 -- native_bf16, artifact artifact--zai-org.glm-5.3-flash-bf16.a6c167b6 @a6c167b62691b2bac901344b65cb651a70f53e43 Metric mean_tokenwise_kld, direction reference_to_candidate, accumulation float64 Estimation surface stack_relation same_stack, head_policy native_head Comparability key cmp--18990ab191ea7a67 Like-for-like predicate comparable: false -- a RECORDED secondary dimension differs across members: pipeline. Equal keys make these rows candidates for comparison, not certified like-for-like; ranking across the differing dimension attributes a lane/pipeline/hardware/scope effect to quantization quality. Machine-readable form with per-dimension values: this key's comparability block in index.json.

What this table is. Every row here shares the comparability key above: the same tokens, the same teacher capture, the same metric and direction, the same estimator precision, the same stack relation and the same head policy. That makes them CANDIDATES for ranking -- the key is a necessary partition, not a certificate. Whether they are also like-for-like on the dimensions the key omits (lane, pipeline, scope coverage, hardware) is what the predicate line above answers.

Rank is not a verdict. The table is sorted by fidelity alone, and fidelity buys bits: a larger, higher-bitrate quant will usually sit above a smaller one, which is not news. Read the Size and Codec columns before reading the order, and compare like against like.

What it is NOT comparable to. The nearest neighbouring groups differ in:

• cmp--f0823827adb15376 (2 rows): panel_id panel--glm53.brandonmusic.final-0000 -> panel--glm53.brandonmusic.final25; reference_id reference--brandonmusic.glm53-bf16-fp32-logits.final-0000 -> reference--malaiwah.glm53-bf16-hf.brandonmusic-final25
• cmp--b55c2d693d127f20 (6 rows): metric_name mean_tokenwise_kld -> mean_of_run_means_tokenwise_kld; accumulation_dtype float64 -> unknown

Those numbers are in this file, under their own headings. Quoting one under the other heading is the mistake this layout exists to prevent: the key is a function of the panel, the teacher, the metric, the direction and the estimator, and the validator recomputes it from those fields rather than trusting the stamped value. What that catches is a row filed under a key its own fields do not produce. It does not catch a number attributed to the wrong panel in the first place -- no offline checker can. That is what the receipt digests on every row are for.

Also, and always: every table for a different model. A KL number is a divergence over one model's own vocabulary against that model's own teacher. It is not a quality score that can be carried between models.

Artifact	Codec	Size	mean_tokenwise_kld (nats)	CI95	Top-1	Runs	Attribution	Receipt
brandonmusic GLM-5.3-Flash tr3 4bpw	exl3-mcg @4	175.6 GB	0.0227508	--	93.84 %	1 run, unevidenced	reported by brandonmusic	receipt
brandonmusic GLM-5.3-Flash NVFP4 runtime build	nvfp4 @4	--	0.0682296	--	91.99 %	1 run, unevidenced	reported by brandonmusic	receipt

The same artifact, measured elsewhere in this file. 2 of the artifacts below also carry a number in another table -- on a different panel, teacher or estimator -- and the widest of those spans 25%. None of the readings is wrong and none is interchangeable with another. Quoting one of them as the number for the artifact, without its table, is the misuse this registry exists to make obvious.

• brandonmusic GLM-5.3-Flash NVFP4 runtime build -- 4 values here, from 0.0547574 to 0.0682296 nats (25% apart). Other tables: cmp--b55c2d693d127f20.
• brandonmusic GLM-5.3-Flash tr3 4bpw -- 3 values here, from 0.0227508 to 0.0249488 nats (10% apart). Other tables: cmp--202b717f3219c414, cmp--2b9c401d13806d7e.

Disclosures for the rows above (6)

• glm53.brandonmusic-4bpw.tp2-runtime.brandonmusic-final-0000 author_reported_only: brandonmusic's custom TP2 runtime on the single qualification window. The receipt notes runtime_raw_decoded_parity_passed false with runtime_rank_output_identical true.
• glm53.brandonmusic-4bpw.tp2-runtime.brandonmusic-final-0000 single_run: One run.
• glm53.brandonmusic-4bpw.tp2-runtime.brandonmusic-final-0000 note: THE PANEL-SCOPE OBJECT LESSON: the same artifact reads 0.022751 here and 0.024555 over the full 25 windows, against the same teacher. A 7% swing from window selection alone.
• glm53.nvfp4-dynamic-scale-control.brandonmusic-final-0000 author_reported_only: brandonmusic's dynamic-scale CONTROL for the v44 NVFP4 row: same window, same teacher, dynamic instead of calibrated power-of-two scales.
• glm53.nvfp4-dynamic-scale-control.brandonmusic-final-0000 single_run: One run.
• glm53.nvfp4-dynamic-scale-control.brandonmusic-final-0000 quality_gate_failed: mean_kld_gate_passed false at threshold 0.06.

Panel: orcarouter MLX evaluation set (undisclosed)

Panel disclosure -- undisclosed_panel: Neither the token set, the window count nor the scored-position total is published. Numbers on this panel can be reported but cannot be compared with anything measured on a known panel -- including other rows for the same model.

Group cmp--492e9b16e8bd6fbd -- 5 rows

Panel panel--orcarouter.undisclosed -- orcarouter MLX evaluation set (undisclosed) -- contexts x -- scored positions = undisclosed scored positions, score_from None sealed: no -- contamination scan: NOT RUN Reference (teacher) reference--orcarouter.glm53-fp8-dequantized.undisclosed -- dequantized_from_quant, artifact artifact--orcarouter.glm-5.3-flash-fp8-dequantized @unpinned revision Metric mean_tokenwise_kld, direction reference_to_candidate, accumulation unknown Estimation surface stack_relation same_stack, head_policy unknown Comparability key cmp--492e9b16e8bd6fbd Like-for-like predicate comparable: unknown -- no recorded difference, but hardware, harness, replay_backend, replay_env, scope, stack are unrecorded for at least one member, so homogeneity cannot be certified. Machine-readable form with per-dimension values: this key's comparability block in index.json.

What this table is. Every row here shares the comparability key above: the same tokens, the same teacher capture, the same metric and direction, the same estimator precision, the same stack relation and the same head policy. That makes them CANDIDATES for ranking -- the key is a necessary partition, not a certificate. Whether they are also like-for-like on the dimensions the key omits (lane, pipeline, scope coverage, hardware) is what the predicate line above answers.

Rank is not a verdict. The table is sorted by fidelity alone, and fidelity buys bits: a larger, higher-bitrate quant will usually sit above a smaller one, which is not news. Read the Size and Codec columns before reading the order, and compare like against like.

What it is NOT comparable to. Every other table in this file: no other group shares this key. That includes every table for a different model -- a KL number is a divergence over one model's own vocabulary against that model's own teacher, never a score that can be carried between models.

Artifact	Codec	Size	mean_tokenwise_kld (nats)	CI95	Top-1	Runs	Attribution	Receipt
orcarouter GLM-5.3-Flash-MLX 6-bit	mlx-affine @6	295.6 GB	0.0063	--	97.76 %	1 run, unevidenced	reported by orcarouter	model_card
orcarouter GLM-5.3-Flash-MLX 4-bit	mlx-affine @4	204.0 GB	0.0131	--	96.13 %	1 run, unevidenced	reported by orcarouter	model_card
orcarouter GLM-5.3-Flash-MLX 3-bit	mlx-affine @3	184.3 GB	0.0421	--	92.06 %	1 run, unevidenced	reported by orcarouter	model_card
orcarouter GLM-5.3-Flash-MLX 2-bit	mlx-affine @2	145.0 GB	0.1647	--	86.56 %	1 run, unevidenced	reported by orcarouter	model_card
orcarouter GLM-5.3-Flash-MLX 2bit-lite	mlx-affine @2	102.5 GB	0.3456	--	77.19 %	1 run, unevidenced	reported by orcarouter	model_card

Disclosures for the rows above (30)

• glm53.orcarouter-mlx-6bit.undisclosed author_reported_only: Reported by orcarouter on their model card. No receipt, no estimator precision, no run count.
• glm53.orcarouter-mlx-6bit.undisclosed different_reference_kind: Measured against the official FP8 release DEQUANTIZED TO BF16, not against a BF16 teacher. This changes the estimand without a guaranteed bias direction. This row's 6-bit 0.0063 is NOT better than the K6 6bpw 0.013723 on brandonmusic's panel -- they are not the same quantity.
• glm53.orcarouter-mlx-6bit.undisclosed undisclosed_panel: Evaluation set not disclosed: no token digest, window count or position total.
• glm53.orcarouter-mlx-6bit.undisclosed subset_of_panel: Panel coverage unknown, so covers_full_panel is false by default.
• glm53.orcarouter-mlx-6bit.undisclosed estimator_unknown: Accumulation precision and head policy are not published.
• glm53.orcarouter-mlx-6bit.undisclosed note: Perplexity reported alongside on the same card: 2.7864 (FP8 reference 2.7797).
• glm53.orcarouter-mlx-4bit.undisclosed author_reported_only: Reported by orcarouter on their model card. No receipt, no estimator precision, no run count.
• glm53.orcarouter-mlx-4bit.undisclosed different_reference_kind: Measured against the official FP8 release DEQUANTIZED TO BF16, not against a BF16 teacher. This changes the estimand without a guaranteed bias direction. This row's 6-bit 0.0063 is NOT better than the K6 6bpw 0.013723 on brandonmusic's panel -- they are not the same quantity.
• glm53.orcarouter-mlx-4bit.undisclosed undisclosed_panel: Evaluation set not disclosed: no token digest, window count or position total.
• glm53.orcarouter-mlx-4bit.undisclosed subset_of_panel: Panel coverage unknown, so covers_full_panel is false by default.
• glm53.orcarouter-mlx-4bit.undisclosed estimator_unknown: Accumulation precision and head policy are not published.
• glm53.orcarouter-mlx-4bit.undisclosed note: Perplexity reported alongside on the same card: 2.862 (FP8 reference 2.7797).
• glm53.orcarouter-mlx-3bit.undisclosed author_reported_only: Reported by orcarouter on their model card. No receipt, no estimator precision, no run count.
• glm53.orcarouter-mlx-3bit.undisclosed different_reference_kind: Measured against the official FP8 release DEQUANTIZED TO BF16, not against a BF16 teacher. This changes the estimand without a guaranteed bias direction. This row's 6-bit 0.0063 is NOT better than the K6 6bpw 0.013723 on brandonmusic's panel -- they are not the same quantity.
• glm53.orcarouter-mlx-3bit.undisclosed undisclosed_panel: Evaluation set not disclosed: no token digest, window count or position total.
• glm53.orcarouter-mlx-3bit.undisclosed subset_of_panel: Panel coverage unknown, so covers_full_panel is false by default.
• glm53.orcarouter-mlx-3bit.undisclosed estimator_unknown: Accumulation precision and head policy are not published.
• glm53.orcarouter-mlx-3bit.undisclosed note: Perplexity reported alongside on the same card: 3.0566 (FP8 reference 2.7797).
• glm53.orcarouter-mlx-2bit.undisclosed author_reported_only: Reported by orcarouter on their model card. No receipt, no estimator precision, no run count.
• glm53.orcarouter-mlx-2bit.undisclosed different_reference_kind: Measured against the official FP8 release DEQUANTIZED TO BF16, not against a BF16 teacher. This changes the estimand without a guaranteed bias direction. This row's 6-bit 0.0063 is NOT better than the K6 6bpw 0.013723 on brandonmusic's panel -- they are not the same quantity.
• glm53.orcarouter-mlx-2bit.undisclosed undisclosed_panel: Evaluation set not disclosed: no token digest, window count or position total.
• glm53.orcarouter-mlx-2bit.undisclosed subset_of_panel: Panel coverage unknown, so covers_full_panel is false by default.
• glm53.orcarouter-mlx-2bit.undisclosed estimator_unknown: Accumulation precision and head policy are not published.
• glm53.orcarouter-mlx-2bit.undisclosed note: Perplexity reported alongside on the same card: 4.3622 (FP8 reference 2.7797).
• glm53.orcarouter-mlx-2bitlite.undisclosed author_reported_only: Reported by orcarouter on their model card. No receipt, no estimator precision, no run count.
• glm53.orcarouter-mlx-2bitlite.undisclosed different_reference_kind: Measured against the official FP8 release DEQUANTIZED TO BF16, not against a BF16 teacher. This changes the estimand without a guaranteed bias direction. This row's 6-bit 0.0063 is NOT better than the K6 6bpw 0.013723 on brandonmusic's panel -- they are not the same quantity.
• glm53.orcarouter-mlx-2bitlite.undisclosed undisclosed_panel: Evaluation set not disclosed: no token digest, window count or position total.
• glm53.orcarouter-mlx-2bitlite.undisclosed subset_of_panel: Panel coverage unknown, so covers_full_panel is false by default.
• glm53.orcarouter-mlx-2bitlite.undisclosed estimator_unknown: Accumulation precision and head policy are not published.
• glm53.orcarouter-mlx-2bitlite.undisclosed note: Perplexity reported alongside on the same card: 6.7018 (FP8 reference 2.7797).

---

Using the data

data/models.jsonl        the upstream models
data/artifacts.jsonl     one concrete weight set at one pinned revision + a STRUCTURED quantization scope
data/panels.jsonl        the token sets, including scored-position policy, sealing and contamination guard
data/references.jsonl    teacher captures: (artifact, panel, stack, precision, head source)
data/pipelines.jsonl     the measuring and producing stacks
data/measurements.jsonl  the rows
index.json               counts, collection digests, and the comparability-key groups as DATA
schema/*.schema.json     JSON Schema draft 2020-12
schema/invariants.json   the machine-readable rules the validator enforces, with severities

Resolver rule: every *_ref is the id of a record in the collection named by the ref's id prefix (model--, artifact--, panel--, reference--, pipeline--, measurement--). That is the only thing a consumer needs to know to join the files.

Query it with one line of jq -- the mission's original complaint, answered:

# every measured quant of GLM-5.3-Flash with its number, panel and who measured it
jq -r 'select(.model_ref=="model--zai-org.glm-5.3-flash")
       | [.metric.value, .artifact_ref, .panel_ref, .provenance.measured_by, .comparability.key]
       | @tsv' data/measurements.jsonl | sort -n

# only rows you may legitimately rank against our K6
jq -r --arg k cmp--202b717f3219c414 'select(.comparability.key==$k)
       | [.metric.value, .artifact_ref] | @tsv' data/measurements.jsonl | sort -n

Tools

python3 tools/registry_validate.py                  # schema + every invariant, offline, no installs
python3 tools/registry_validate.py --strict --json  # CI mode
python3 tools/registry_validate.py --explain <id> [--against <id>]
python3 tools/registry_render.py [--check]          # regenerate / verify README tables + index.json
python3 tools/registry_add.py from-receipt --receipt R --artifact A --panel P ...
python3 tools/seed_registry.py --check              # the seeded rows are regenerable (see the note below)
make check                                          # validate + render --check + fixtures

What seed_registry.py --check does and does not prove. The 37 Qwen3.8-27B rows are read live out of their receipt files on every run — the seeder refuses to build if a receipt is missing — so --check genuinely re-derives those values from receipts and byte-compares them. The 20 GLM-5.3-Flash rows are transcribed literals: their receipts live on the Hub and in third-party repositories, and this tooling is offline by contract, so for those rows --check proves that data/ matches seed_registry.py, not that seed_registry.py matches the receipt. Each of those rows records its receipt's sha256, so the binding is checkable by hand: fetch the uri, hash it, and compare the value at the field_provenance pointer. All 20 were checked that way on 2026-08-28 and all 20 matched at full float64. Nothing in CI rechecks it, because nothing in CI is allowed to reach the network.

Both tools run on a stock interpreter with no network and no pip: tools/_minischema.py is a vendored draft-2020-12 validator covering exactly the keyword subset these schemas use, and it raises on any keyword it does not implement rather than silently ignoring it. When the real jsonschema library is importable, --jsonschema-lib both runs both and the CI job fails if their verdicts differ, so the vendored one cannot quietly drift.

Credit

The artifacts and the numbers in this registry mostly belong to other people. Specifically:

• brandonmusic built the sealed GLM-5.3-Flash token panel, captured and published the fp32 BF16 teacher logits that four of our own numbers are measured against, produced the tr3-4bpw checkpoint, and measured and published the 4bpw and runtime-image rows on his own stack. The panel and the teacher are his work; we are guests on them.
• 0xSero produced the GLM-5.3-Flash EXL3 Q4 (Dione) release. The Q4 number here is ours, the artifact is theirs. local-ai-registry is also theirs, and this registry is shaped to interoperate with it.
• orcarouter (Continuum AI Corp) produced the GLM-5.3-Flash MLX builds and reported their own fidelity numbers, which are included here as their measurements against their reference, quarantined from ours rather than merged into them.
• turboderp wrote exllamav3, without which most of the EXL3 artifacts in this registry would not exist, and published the Qwen3.8-27B exl3 branches measured here.
• Z.ai (zai-org) published GLM-5.3-Flash and its official FP8 release. Qwen (Alibaba) published Qwen3.8-27B and its FP8 release. unsloth, gittensor-model-hub and the authors of the AWQ-INT4 and MTP-NVFP4 builds produced artifacts we measured.

Where an upstream author's identity could not be established from a receipt, this registry records repository: null and says so, rather than asserting a repo id it cannot back up.