• Status: FROZEN at the commit that introduces this file. Amendments ship as later numbered documents, never as edits here.
• Date: 2026-08-10. Revised same day after an adversarial soundness review (18 findings applied before the freeze).
• Relationship to PREREG-002: that document stays frozen and governs the executed DS-002 / DS-002b runs. Per ERRATA-POLICY's pointer-line rule, PREREG-002's preamble (outside its frozen sections 2-7) gains one clearly marked non-normative pointer to this document, shipped in a commit after this file's freeze commit.
• Method by reference: buffer construction conventions, the match predicate, and the reproduction path are PREREG-002 §4 and §5 and KMS-002 §3, unchanged. This document registers populations, variants, a new control, and verdict rules.

1. Why these runs

The 2026-08-10 audit established, from the committed datasets:

1. The frozen candidate classifier (internal/certclass/models.go) omits layer families that PREREG-002 §3 declares mineable. The known consequence: 10 corpus blocks declaring (n 2,048, k 1,024), which is the Qwen3-30B attention o_proj shard at tp4, were never scanned against the tensor family their declared shape implies (as control members they were hashed only against Llama buffers), and all 10 sit inside DS-002's 1,000-block negative control. One previously excluded cell was chased outside this document: the Gemma attention-variant o at k 16,384 produced a byte-exact match at h68,332, recorded as a dated erratum in the roadmap; that match is not reproducible from any committed results file, and the sweep that found it is committed as its own dataset before R1 results are reported.
2. DS-002's negative control does not satisfy the null PREREG-002 §1 stated for it ("commit to differently-shaped operands"): 10 of its 1,000 members declare a published-checkpoint shape, its diversity is exactly 7 distinct (n, k) with 824 of 1,000 at one shape, and the same 1,000 blocks were reused across all three runs.
3. The unmatched residue of the matched era clusters in ways that suggest convention rather than weight differences: 162 of the 163 unmatched varying-m candidates are o-family shapes, and 1,904 of the 2,053 unmatched constant-m 8,192 candidates are the 70B gate_up tp2 shape in a family whose standard layout is match-validated.

2. Registered runs and populations

All populations are defined over the committed, frozen datasets by exact filter, so every population is reproducible by command before any new hashing starts. Notation: matched is the set of heights with match == true in any DS-002 or DS-002b results file; extract is DS-002's extract.jsonl.gz; the span is DS-001's, heights 1 to 96,405 (DS-002's frozen candidate tip is the later 96,774; the difference is the classification snapshot, stated here so no reader reconciles them silently).

R1, coverage closure. Enumerate every (model, layer family, tensor-parallel degree, shard) cell implied by PREREG-002 §3's mineable table that the frozen classifier omits. Dims are read per tensor from each pinned checkpoint's safetensors headers, never from config defaults, and per-layer dim variants are enumerated as distinct cells (the h68,332 class, an attention-variant dim invisible in the config, is exactly what config-derived enumeration re-misses). The enumeration also includes, clearly labeled exploratory, the alternative-sharding cells for o (column-parallel shard shapes) for EVERY pinned checkpoint, Llama included, since a miner using that convention would declare those shapes rather than the row-parallel ones. The buffer for an exploratory column-parallel cell at rank r of tp is the contiguous row block [r n/tp, (r+1) n/tp) of the row-major (n, k) tensor. Procedure (executor: scripts/prereg3-cells.py, named here so the freeze is auditable): expand each family to sharded (n, k) shapes under the PREREG-002 §5 conventions with tp in {1, 2, 4, 8}; all dims come from safetensors headers, and the kv-head replication boundary comes from the config, which is a sharding-rule parameter rather than a dim; drop cells violating the frozen shape floors (k at least 1,024 and n at least 256, certclass MinK/MinN, with the sharded dim divisible by tp; the enumerator additionally requires k divisible by 64, which drops no cell in the pinned checkpoints); subtract cells the classifier or the executed scans already carry, supplied as a mandatory input that is exactly the cells recorded in the DS-002 and DS-002b MANIFESTs (extraction manifests for not-yet-hashed buffers are excluded, as is the C4 sweep cell) and is committed alongside cells.json, making it identical to the executed-scan set R5's eligibility names; emit every remaining cell as its own row (no shape-level deduplication; sampled height lists are labeled as samples), with counts, as DS-006-coverage-closure-v1/cells.json, committed before any hashing; scan every nonzero cell. A family whose headers are not I8 is not enumerated, and that outcome is reported as a correction to PREREG-002 §3's table, never as a silently empty enumeration. Known in advance: the Qwen3-30B attention cells (o full (2,048, 4,096); fused qkv (5,120, 2,048); per-tp shards accordingly), of which the o-at-tp4 cell has the 10 blocks of §1.1; and the exploratory 70B o column-parallel tp2 cell (4,096, 8,192), which holds 371 DS-001 blocks, all OFFICIAL_PATTERN_NONMODEL at constant m 4,096, none in DS-002's control, never scanned. The Qwen3 checkpoint is pinned at revision 751fd1052d4b25faf16072173493b15097a12901.

R2, o byte-order probe. Population, by exact filter: extract records with control == false, candidates nonempty, height not in matched, m not in {8,192, 32,768}, and candidate family o (162 blocks; the 163rd varying-m residual, h48,687, is gate_up tp2 and is probed in R3). One variant, fixed here: V1, column-major byte order of the canonical buffer (the bytes of the transposed matrix laid out in the declared (n, k) shape; a framework storing the tensor transposed would commit exactly these bytes). V1 preserves the declared shape for every population member, square or not. Built-in positive control: 20 blocks drawn from the 144 matched o blocks (120 70B tp1, 21 70B tp2, 1 8B tp1, 2 Gemma tp2), hashed under the standard convention (must match) and under V1 (must not match); any deviation voids the variant run. Reversed-shard-order and column-parallel buffer variants are deliberately absent: every shard index of every searched family was already hashed in DS-002/DS-002b, so shard-order permutations cannot produce a new match, and column-parallel sharding implies different declared shapes, which R1's exploratory cells cover instead.

R3, gate_up tp2 permutation sweep. Population, by exact filter: extract records with control == false, height not in matched, candidates == ["Llama-3.3-70B/gate_up_fused/tp2"], and (m == 8192 or height == 48687): 1,905 blocks, the 1,904 constant-m members plus the one varying-m block of the same shape, so no residual is orphaned between R2 and R3. Permutation set, fixed here: P1 up|gate concatenation order (standard is gate|up); P2 slicing the already-fused tensor per rank (standard, per PREREG-002 §5, shards each sub-projection first and concatenates per rank; P2 is the historically plausible defective layout). Positive control: 20 blocks drawn from the 80 matched gate_up tp2 blocks, same must-match / must-not-match rule as R2.

R4, third-party quantization scan. Population: the 2,216 unmatched candidates outside the m 32,768 stratum: R2's 162, R3's 1,905, plus the remaining 149 unmatched m 8,192 candidates (137 70B o tp1, 11 8B gate_up tp2, 1 70B o tp2). Candidates: published checkpoints of the same architectures whose quantization stores int8 weight tensors for the relevant families, pinned by revision at execution. Two verified eligible before this freeze: RedHatAI/Llama-3.3-70B-Instruct-quantized.w8a8 and CalamitousFelicitousness/Llama-3.3-70B-Instruct-W8A8-INT8; 8B candidates are screened by the same rule (config-declared int8 symmetric weight storage) before download. Buffers follow PREREG-002 §5 applied to the third-party checkpoint. Self-requantization from base weights is out of scope: checked 2026-08-10, the published source tree carries only a runtime GPU activation-quantization kernel, no offline checkpoint recipe, and the base checkpoints are gated.

R5, control redraw. A new negative control replacing DS-002's for all runs registered here: per era stratum (pre-boundary h1-54,972; post-boundary pre-MoE 54,973-71,934; MoE window 71,935-91,629; dense-only 91,630-96,250; rank-penalty 96,251-96,405), draw 200 blocks, or every eligible block where a stratum holds fewer than 200 (the rank-penalty stratum holds 155 blocks in total before exclusions); the realized per-stratum n is reported. Eligibility: the block's declared (n, k) appears neither in the committed cells.json (R1's enumerated list, including the exploratory cells) nor among the cells recorded in the DS-002 and DS-002b MANIFESTs; the R5 draw therefore runs only after the cells.json commit. The draw's shape composition is reported, not forced, and this null differs from PREREG-002 §1's by design: eligibility is shape-only rather than class-restricted, so the eligible pool includes model-shaped and official-consistent blocks whose declared shapes match no searched cell, which is sufficient for the canary purpose (an unsearched shape cannot match) and is stated here so nobody reconciles the two nulls silently. The spike-in positive control is rerun in a scratch copy of the R5 extract, as in OBS-005 §8, so the recorded run is unaffected.

Randomness: seed = 0x50524733 + offset, with offset 2 for R2's positive-control draw, 3 for R3's, and 5 + i for R5 stratum i (i = 0..4 in the order listed above). Each draw is random.Random(seed).sample(heights_sorted_ascending, k) with k = min(target, len(eligible)), under CPython 3.11 or later. A run's variants share one draw (R2's V1 uses the offset-2 draw; R3's P1 and P2 share the offset-3 draw).

3. Predictions, stated before running

• R1: no directional prediction for the 10 Qwen3 o-tp4 blocks; both outcomes are informative (a match triggers the §4 contamination branch; a zero bounds the cell, layout-conditionally per §6). For all other enumerated cells the expectation from the corpus record is zero or near-zero matches, with one exception: the Gemma attention-variant o cell (5,376, 16,384), whose one block h68,332 is a known match being re-derived into a committed results file; a non-match there falsifies the roadmap erratum and is reported as exactly that.
• R2/R3: exploratory. A hit is a layout-convention discovery about buffer construction, never a new-weights finding, and reclassifies the affected non-matches; the correction ships as a new numbered note.
• R4: expectation is zero (different quantization grids should not produce byte-identical int8 codes); a hit is a new attested-species finding.
• R5: expectation is zero matches; see §4 for the rule.

4. Verdict rules, including the pre-decided contamination branch

PREREG-002 §6 rules "any control match: implementation defect, the run is void". That rule presumed a correctly specified control. For the runs here:

1. A byte-exact match on a member of DS-002's ORIGINAL control under R1 is control contamination: a finding about the control's specification, not a scanner defect and not a voiding event. It is verified through the independent raw-bytes reproduction path before being reported. What it corrects is precisely stated: DS-002's "0 matches in 840,000 control pairs" remains literally true (those pairs never included a Qwen3 attention buffer); what falls is PREREG-002 §1's specification premise ("control blocks commit to differently-shaped operands") and the canary interpretation resting on it. The correction ships as a dated errata block in OBS-005, and the external deposit gains a new version naming the errata, per ERRATA-POLICY. A new numbered note is written only if a central finding of OBS-005 reverses, which this branch by itself does not.
2. An R1 match on a block that is NOT an original-control member is an ordinary positive result under PREREG-002 §6's first row: attested model-weight mining, extended to a new cell.
3. A match inside the NEW R5 control is an implementation defect and voids the affected run, exactly as PREREG-002 §6 intended, because R5's null is specified against the full searched cell set.
4. An R4 match is a new attested-species finding, reproduced per §5 and shipped as its own numbered note; it voids nothing.
5. R2/R3 variant hits reclassify; they do not void. A variant that fails its built-in positive control is itself void (wrong variant implementation), reported as such.
6. Every zero is reported with its exact cell list, buffer set, revision pins, and block population, never as proof of absence.
7. Ordering: the OBS-005 §10 dated errata block recording the h68,332 correction lands before any PREREG-003 run reports anything about the extinction tail.

5. Artifacts and freeze mechanics

• Results are emitted append-only into docs/research/datasets/DS-006-coverage-closure-v1/ (R1-R4, including cells.json) and docs/research/datasets/DS-007-stratified-control-v1/ (R5), each with MANIFEST sha256 lists, the emitting tool identity, and this document's commit hash recorded in the summary as prereg3Commit.
• The commit that adds this file is its freeze point. The PREREG-002 pointer line and the cells.json commit both come after it.
• Every new match is independently reproduced from raw Blockbook bytes (scripts/verify-hashb-match.py) before it is reported anywhere.

6. Known limitations, stated in advance

• R1 dims for Qwen3 attention derive from safetensors headers and fused-layer conventions source-derived from the mining stack; no Qwen3 attention layout is match-validated in advance. A zero in those cells is layout-conditional, exactly as PREREG-002 §5 warns for unvalidated layouts, and is reported with that caveat.
• R4 covers only checkpoints that publish int8 weight storage; GPTQ packed formats and FP8 storage are out of scope by construction.
• Nothing here changes the structural ceiling: private, requantized, training, or synthetic operands remain mutually indistinguishable on chain, and no zero below supports a claim about useful work in general.