Dated observation note, internal research record. Recorded 2026-08-10. Status: necessary-condition screen, exploratory, not pre-registered. Status revision, 2026-08-11: "internal research record" is superseded as to publication only. ADR-0014 re-scoped publication to the site package and this note ships inside it; the exploratory, not-pre-registered status and §2's coverage bounds are unchanged. Covers roadmap C4 (the "ended vs moved" question; PLAN-002 §9.3 item 5, STATUS-2026-08-09 §5 item 6). Verdict: ENDED within §2's bounds, and the verdict criterion is computed in code, never eyeballed: the tool itself prints a per-(model, era) PASS/FAIL with reasons and the summary line quoted in §4. Everything regenerates from two frozen committed inputs via scripts/shape-sweep-nonpearl.py plus the companion scripts/nonllm-shape-check.py; both tools' outputs are committed verbatim as DS-008, and every number in this note is read from those committed artifacts or from a command stated inline.

Sources: frozen DS-001 (declared per-block m, n, k, class; non-empty blocks h1-96,405, n = 96,405, genesis EMPTY excluded); shape-dictionary-v1.json (generated 2026-08-09, config-derived at pinned revisions: 36 checkpoints, of which 4 pearl-reference and 32 non-pearl targets, 603 grid entries); the committed DS-002/DS-002b results for the era-boundary verification (§3); Keshi's internal protocol notes (Certificate internals) for the consensus shape floor; nine further checkpoints pinned by revision sha inside nonllm-shape-check.py (§6). Unit of analysis throughout: blocks, one certificate per block. Checkpoint names below are subject-matter identifiers.

Label discipline, per metrics.md §4: only a byte-exact hash_b match is "attested model-weight mining", and nothing is ever "certified inference". A shape match is strictly weaker than either: it is a necessary condition, never sufficient, and this note contains no new weight attestations at all.

1. Question

OBS-005/DS-002 established that attested model-weight mining was a launch-week phenomenon of a handful of operators. Its scan could only test the four pearl-ai reference checkpoints, so the standing "~96% of current blocks declare dims matching no reference model" (STATUS-2026-08-09 §2, current-window blocks) meant "not shaped like those four". Two hypotheses fit that record: real-weight mining ENDED, or it MOVED to weights the scan never searched. This sweep bounds the one branch of "moved" that declared data can reach: moved to a KNOWN PUBLIC checkpoint whose linear-layer geometry is in the dictionary. It reads only declared certificate dims, which are free choices; consensus never checks them against any model (pearl-notes.md §Certificate internals).

2. What the verdict can and cannot mean

Stated before the result so the result cannot be over-read.

1. A shape match is necessary for real mining of a dictionary checkpoint under the standard serving layouts, and nothing more. Only a hash_b weight match proves committed operands (OBS-005 §0.2). A shape match here would have justified a targeted weight scan, not a finding.
2. A non-match proves nothing about private models, fine-tunes, requantized variants, training runs, or synthetic work. Those are mutually indistinguishable on chain by construction (OBS-005 §0.1, the verifiability ceiling), and §7 restates what that does to the "moved" hypothesis.
3. The screen assumes whole-layer tiles under the dictionary's fusion and tensor-parallel sharding conventions. A miner slicing nonstandard tiles from a known checkpoint would evade it (inferred bound; the same genus as PREREG-003's byte-order probes).
4. Coverage is finite and partially shadowed (§5): "any known public checkpoint" means the 36 dictionary checkpoints minus the shadowing losses, nothing broader.

3. Method

Dictionary: for each checkpoint, every mineable linear layer expanded in the fused serving layout (qkv_fused, o, gate_up_fused, down, MoE variants with their (experts, topK) trailer) across tp in {1, 2, 4, 8}, keeping entries with n >= 256 and k >= 1,024 (the dictionary floors; minN/minK in the dictionary header). The k floor coincides with the consensus floor; the n floor is the vLLM miner engagement floor (certclass.go MinN), so sub-256-n declarations are consensus-legal but out of dictionary scope. Dims are config-derived at pinned revisions; per-tensor variants invisible in a config are a known blind-spot class. The one known instance, the Gemma-4-31B full-attention o variant (5376, 16384) found at h68,332, is present in this dictionary build (verified by command against the committed dictionary), and PREREG-003 R1 closes the class for the pearl checkpoints by reading safetensors headers.

Matching mirrors the frozen T2 semantics: exact ordered (n, k) equality; dense rows require an empty MoE trailer, expert rows the exact trailer (trailer-inconsistent hits are reported separately, never counted). Any pair a pearl-reference checkpoint can produce (57 distinct pairs) is excluded from the non-pearl increment wherever it also appears, leaving 241 distinct non-pearl pairs. Cross-checks from the run: 0 DS-001 T2-matched blocks fall outside the pearl exclusion set (the wanted value is 0), and 11 blocks are pearl-shaped here without being DS-001 T2-matched (expected: config-derived rows the frozen classifier lacked, h68,332 among them).

Era split: the attested-mining extinction boundary h54,972, then the three fork heights 71,935 / 91,630 / 96,251. The boundary's meaning was re-verified by command over the committed DS-002, DS-002b-gemma and DS-002b-o-tp48 results (1,109 matched rows, 1,109 distinct heights): the last matched height at or below 54,972 is exactly 54,972, and only three matched heights sit above it (56,217 and 62,319 in the Gemma results, 60,881 in DS-002), plus the h68,332 match recorded in the roadmap §8.4 erratum note (source: ROADMAP.md, "the last attested model-weight block moves ... to h68,332"). All four stragglers declare pearl-reference pairs (verified by command against DS-001 and the dictionary), so they sit in S2's pearl column below and cannot leak into the non-pearl increment. "Post-extinction" in this note therefore means "after the contiguous attested era", with exactly these four known attested stragglers inside S2.

Specificity bar, computed in code per (model, era): PASS only if the model matched at least 2 distinct (n, k) pairs from its grid in that era, or exactly one pair that is simultaneously dictionary-unique (no other target shares it), has neither dim a power of two, and shows varying m (more than one m value, and not the constant m == n signature). Everything else is compatible with coincidence. The constant m == n reading as a dims-grinder rather than batched inference is inferred, not proven; the bar only demands that a "moved" candidate look unlike one.

4. Result

The tool's per-era table, verbatim from the committed sweep-report.txt (denominators in the legend):

Per-era table. Denominators: blocks = all non-empty DS-001 blocks in the segment; p% over blocks; m%np and p2%np over NON-pearl-shaped blocks in the segment; pow2n counts non-pearl blocks whose declared n is a power of two; none = matches no dictionary shape at all.
segment                                   blocks   pearl     p%  match     m%   m%np  trlr   pow2n  p2%np    none
-----------------------------------------------------------------------------------------------------------------
S1 attested era        h1-54,972           54972   50171  91.27    371   0.67   7.73     0    4592  95.65    4430
S2 post-extinction     h54,973-71,934      16962    2508  14.79      0   0.00   0.00     0   13606  94.13   14454
S3 moe-window          h71,935-91,629      19695     808   4.10     71   0.36   0.38     0   14881  78.79   18816
S4 dense-only          h91,630-96,250       4621     226   4.89      9   0.19   0.20     0    2906  66.12    4386
S5 rank-penalty        h96,251-              155       3   1.94      0   0.00   0.00     0      71  46.71     152

Reading it, all figures from the committed run:

• S2, the cleanest window (16,962 blocks, of which 14,454 non-pearl-shaped): zero dictionary hits of any kind.
• S5 (155 blocks, 152 non-pearl-shaped): zero hits.
• S3 + S4 carry 80 raw hits out of 24,316 blocks in those two segments (23,282 of them non-pearl-shaped); across all of S2-S5 that is 80 of 41,433 non-empty post-extinction blocks. Every one of the 80 is a coincidence by the computed bar; they reduce to two shapes:
  • (5120, 4096), 68 blocks, h75,253-93,837: shared by 6 grid cells of 5 different models (o and down cells of Mistral-Small-24B, Llama-4-Scout-17B-16E, Mistral-Nemo-12B, Qwen3-32B, DeepSeek-V2), one dim a power of two, and every block at constant m = 5120 = n.
  • (12288, 4096), 12 blocks, h75,309-93,087: Qwen3-8B gate_up tp2, one dim a power of two, every block at constant m = 12288 = n.
• S1's 371 hits are (4096, 8192), all at constant m = 4096 = n, both dims powers of two, nominally Qwen3-235B-A22B o at tp1. Same verdict: coincidence-compatible, and inside the attested era anyway.
• Trailer-inconsistent shape hits: 0 in every segment. The third-party pearl-quantized Qwen3.6-27B-heretic checkpoint: 0 blocks in any era.
• All-era raw hits: 451 blocks; every (model, era) candidate fails the bar, each with its printed reason (constant m == n, power-of-two dims, or a shape shared across models).

The tool's committed summary line:

Bar summary: 0 (model, era) pairs PASS the specificity bar across all eras. A zero here is the 'ENDED (no move to a known public checkpoint)' verdict.

ENDED, exactly this much: in 41,433 non-empty post-extinction blocks, no model in the dictionary shows a signature that survives the specificity bar; the clean windows S2 and S5 show zero raw hits over 14,454 and 152 non-pearl-shaped blocks respectively. Useful-work mining of a known public checkpoint's standard serving geometry did not resume under a different model family within the dictionary's coverage (§5). The verdict says nothing about weights the chain cannot verify (§7), and it is falsifiable by future blocks: two post-extinction certificates matching distinct cells of one target's grid, or a single certificate on a dictionary-unique cell with neither dim a power of two and m unequal to n, would flip the bar to PASS.

5. Coverage and shadowing caveats

The pearl exclusion protects the increment from false attribution and costs coverage; both losses are printed by the run.

• Fully shadowed targets, 5 of the 32 non-pearl targets: every grid cell equals a pearl-reference pair, so the sweep has zero discriminating power for Llama-3.1-8B-base, Llama-3.1-70B-base, Mistral-7B-v0.3, Mixtral-8x7B, and Qwen3-30B-A3B-base (dimension twins of the pearl set). Mining of these is indistinguishable from pearl-reference-shaped traffic by geometry alone.
• Partially shadowed targets, 16 of the 32 (shadowed cells / grid cells): DeepSeek-V2 2/29, DeepSeek-V2-Lite 5/26, DeepSeek-V3 4/30, GLM-4.5-Air 3/26, Gemma-2-9B 3/15, Gemma-3-27B 14/15, Kimi-K2 4/30, Llama-3.2-3B 1/14, Llama-4-Scout-17B-16E 4/23, Mistral-Nemo-12B 1/15, Mistral-Small-24B 2/15, Mixtral-8x22B 1/15, Qwen2.5-72B 8/16, Qwen3-235B-A22B 6/13, Qwen3-32B 2/16, Qwen3-8B 9/15. Gemma-3-27B is the worst case, screened at 1 of its 15 cells.
• Honest coverage statement: the non-pearl screen genuinely covers 27 of its 32 target checkpoints, and of those 27, 11 at their full grid and 16 only partially. Counting the 4 pearl-reference checkpoints, which the table tracks in the pearl column rather than the increment (as a set; their shared pairs are not attributable to one of the four), 31 of the 36 dictionary checkpoints remain geometrically trackable. Shorthand: 31 of 36, partially.
• Dictionary floors: entries need n >= 256 and k >= 1,024 over tp in {1, 2, 4, 8}, so sub-floor cells of covered checkpoints are absent by construction (this is why per-target grids differ in size, e.g. Llama-3.2-3B at 14 cells).
• Unresolved dictionary targets in this build: 0 (the tool prints any).

6. The non-LLM screen: vision, audio, embedding

The dictionary is LLM-only, so a companion screen (scripts/nonllm-shape-check.py, committed output nonllm.json / nonllm-report.txt in DS-008) tests the same necessary condition for nine popular non-LLM checkpoints. Dims are read from each pinned checkpoint's safetensors headers over HTTP range requests, never from config defaults, across all layers; expansion uses the same fused serving conventions at tp = 1 only, so the screen bounds unsharded serving geometry and nothing else. Checkpoints (pinned by revision sha in the script and recorded in nonllm.json): CLIP-ViT-L-14, CLIP-ViT-H-14, SigLIP-So400m, DINOv2-giant, InternViT-6B (vision), Whisper-large-v3 (audio), BERT-large, BGE-large-en-v1.5, BERT-base (embedding).

The nine checkpoints yield 24 distinct (n, k) hypotheses; the committed run classifies them before counting:

• 20 of the 24 pairs are protocol-legal hypotheses; zero of the 96,405 non-empty DS-001 blocks declare any of them (and the 4 excluded pairs also appear in zero blocks, as consensus requires).
• The 2 pearl-shadowed pairs, (4096, 1024) and (1536, 4096), equal pearl-reference pairs (Llama-3.1-8B o tp4 and qkv_fused tp4), so they would have had zero discriminating power even with hits; they happen to have none.
• Consensus k-floor caveat, binding for interpretation: the verifier enforces k >= 1,024 and k % 64 == 0 (zk-pow sanity_checks.rs:24-55, cited via pearl-notes.md §Certificate internals), so 4 hypothesis pairs cannot appear in any certificate: the k = 768 geometries of BERT-base and the CLIP-ViT-L-14 text tower (3 pairs), and SigLIP-So400m's mlp_out (1152, 4304) with k % 64 == 16. The absence of small vision and text encoders below the floor is therefore partly a protocol artifact, not evidence about miner behavior; for the 768-class the exclusion is partial, since its mlp output pair (768, 3072) is legal, was tested, and is also absent. Geometry that clears the floor (all 20 legal pairs, from the 768-class mlp output through every class at hidden size 1,024 and up) is genuinely absent from all 96,405 blocks' declared dims.
• Bounds: nine checkpoints, fused layout, tp = 1. This is a spot screen of the most-deployed encoder geometries, not a census of non-LLM models; sharded or otherwise nonstandard tilings of these checkpoints are untested (§2, item 3).

7. What this does not establish

• No statement about useful work in general. Per the OBS-005 §0.1 ceiling, job_key is per-block, so only public static checkpoints are externally verifiable; private models, customer fine-tunes, locally requantized variants, evolving training weights and synthetic noise remain mutually indistinguishable on chain. "Moved to weights the chain cannot verify" is untestable by design, and this sweep neither supports nor damages it.
• Not proof of synthetic work. The 37,888 post-extinction non-pearl-shaped blocks match no dictionary shape (or only coincidence-compatible ones), which bounds what they are not; it does not identify what they are.
• Not a weight attestation. No hash_b was computed here; a shape match, had one survived the bar, would itself have proven nothing without a subsequent scan.
• Not proof of absence beyond the search space: 36 LLM checkpoints minus the §5 shadowing losses, 9 non-LLM checkpoints, standard fused layouts, tp grid {1, 2, 4, 8} (LLM) and tp = 1 (non-LLM), floors n >= 256 / k >= 1,024. A checkpoint outside the dictionary, or a nonstandard tiling of one inside it, is not excluded.

8. Reproduction

From the repo root; both tools are byte-deterministic given the frozen inputs (re-runs this session produced byte-identical JSON and reports):

python3 scripts/shape-sweep-nonpearl.py --json /tmp/sweep.json
diff /tmp/sweep.json docs/research/datasets/DS-008-shape-sweep-v1/sweep.json

python3 scripts/nonllm-shape-check.py --json /tmp/nonllm.json
diff /tmp/nonllm.json docs/research/datasets/DS-008-shape-sweep-v1/nonllm.json

The non-LLM tool fetches safetensors headers for the pinned revisions over the network; everything else is offline against the committed inputs. --census 7d on the sweep adds a live-API cross-check, labeled LIVE and NOT frozen, and is not part of DS-008. File hashes: DS-008 MANIFEST. The dataset is append-only; corrections ship as a new version directory with an errata note, per ERRATA-POLICY.md.