Dated observation note, 2026-08-06. This note fills the pre-written interpretation table of PREREG-001, as frozen for this run at 11a517a02cedfb97360da95bfaf8e4f114374dfd, with the measured numbers from dataset DS-001: 96,406 canonical certificates at heights 0–96,405 (the run pinned --to 96405), decoder v1. Nothing in §§1–4 below is an interpretation not written into the prereg before the run; anything else is confined to §5 "Exploratory (post-hoc)".

That commit is what DS-001 was produced under and what its manifest records. The document has since taken a dated typography erratum (2026-08-07) that changed no hypothesis, test, threshold or table value; the erratum itself carries the command to diff the two versions. Read the hash here as the provenance of this run, not as a claim about the document's current bytes.

Binding language (PREREG-001 §0): every statement is shape consistency. No result here proves inference occurred, that a model was served, or that any computation was fresh. Per-block accusations are not made.

1. Class distribution

Per era (activation-at-height strata):

Official-pattern traffic (T1-pass, 23,823 blocks = 24.7% of the chain) is concentrated in the pre-MoE era; from the MoE fork onward the chain is ≥92% CUSTOM.

2. T1 (hard falsifier): no labeled pool is model-consistent, ≥99.5% provably not the unmodified stack

T1's binding asymmetry (PREREG-001 §3): failure proves a block was not produced by the unmodified published stack; a pass proves nothing (env overrides). So T1 licenses two statements about the labeled pools, one hard, one about model-consistency:

Two facts, stated precisely:

1. ≥99.5% of every labeled pool's blocks are provably not the unmodified published stack (T1-fail; ~29,100 attributed blocks, only 64 T1-pass).
2. Zero labeled-pool blocks are model-consistent (OFFICIAL_CONSISTENT). The 64 T1-pass labeled blocks are all OFFICIAL_PATTERN_NONMODEL (official bytes, non-model dims): Pearl Fortune 40, Kryptex 23, PearlHash 1. Those 64 cannot be falsified by T1 (a pass proves nothing), so statement 1 is a near-universal, not a clean universal.

Era-confound checked and refuted. This is not an artifact of pools appearing late: in the pre-MoE era, where unattributed traffic is 30% OFFICIAL_CONSISTENT, PearlHash mined 4,313 blocks (0 OFFICIAL_CONSISTENT), Pearl Fortune 3,314 (0), Kryptex 230 (0). Within the same era and the same consensus rules, labeled pools produce essentially no official-fragment blocks while unattributed miners produce thousands. This is a pool-level difference, not an era one.

(For the relationship to arXiv:2606.04819's pool-binary claim, a non-pre-registered reading, see §5.)

3. T3 (mixture): pre-registered verdicts

Only three strata reach the N ≥ 200 T1-base floor, all unattributed:

Every other stratum is INSUFFICIENT_SAMPLE. No stratum × hypothesis cell reached MIXTURE_CONSISTENT or INDETERMINATE.

Pre-registered interpretation (PREREG-001 §6, outcome row "both forms > 0.25 for all H", copied verbatim): "Mixture inconsistent with every pre-registered model/tp; consistent with synthetic shape selection or an unpublished model. Not proof of synthetic mining."

So even the official-pattern, 70B-layer-shaped traffic does not appear in the MAC-share proportions a served model predicts. Both strata that clear the matched-count floor (pre-moe, moe-window) fail the primary and the conditional form; dense-only fails the primary form (its conditional is INSUFFICIENT_SAMPLE, matched 193 < 200).

4. T4 (m-variability): pre-registered flags

The dominant m is 32,768, above vLLM's typical max_num_batched_tokens (~16k), in 82–100% of these strata's model-shaped blocks. Per the pre-written caveat (PREREG-001 §7), dominant m > 16,384 is "stronger but still configuration-dependent" (the batch bound is operator-set), and constant power-of-two m at a chunk size is weak on its own. The prereg does not license calling this synthetic; it is a flag, not a verdict.

5. Exploratory (post-hoc): the mechanism behind the inconsistency

Not pre-registered; presented as characterization, not as a graded finding.

Why is the largest official-pattern stratum MIXTURE_INCONSISTENT despite 71.3% of its T1 blocks matching a 70B layer pair? The (k, n) distribution is a near point-mass on the single highest-MAC layer:

A real forward pass emits all four layers in MAC proportion; this traffic is overwhelmingly the one shape with the largest n·k (gate/up, at tp1 and tp2 = 78.5% combined) while the down-projection (the second-largest predicted layer, 27.5%) is entirely absent. That is the shape a declared-MAC- maximizing miner would select, and it is what "synthetic shape selection" (§3's pre-registered interpretation) looks like mechanically. It does not prove synthetic mining; a partial or non-standard serving configuration is not excluded. But it is why the mixture test rejects every served-model hypothesis.

Erratum 2026-08-07: a simpler mechanism, and it qualifies §3. The published checkpoints' own quantization_config and safetensors headers (verified directly: PREREG-002 §3) show that down_proj is quantized FP8 (F8_E4M3), not INT7, in every Llama checkpoint: Llama-3.3-70B layer 50's mlp.down_proj.weight is F8_E4M3 [8192, 28672] while gate_proj/up_proj are I8 [28672, 8192]. An FP8 layer does not flow through NoisyGEMM and therefore cannot produce an INT7 certificate at all. The 0.0%-vs-27.5% gap above is not miner behaviour; it is a property of the published model. The same source shows only half of each model's qkv layers are INT7 (70B: layers 40–79; 8B: 16–31), so the qkv share was also overstated.

What this changes. PREREG-001's frozen Llama MAC-share vectors (gate_up 0.549 / down 0.275 / qkv 0.098 / o 0.078 for the 70B) do not describe what the published stack can mine, so §3's MIXTURE_INCONSISTENT verdicts are substantially an artifact of a mis-specified expected distribution and must not be read as evidence of synthetic shape selection. The §5 paragraph above stands only as the description of an observed distribution, not as a mechanism.

What this does not change. §2's T1 result is untouched: T1 is a byte-exact test on rank and tile patterns and never used the layer tables. "No labeled pool is model-consistent, and ≥99.5% of every labeled pool's blocks are provably not the unmodified published stack" stands as reported, as does the era-confound refutation.

PREREG-001 is frozen and is not edited. A corrected mixture test, if pursued, ships as its own pre-registration; PREREG-002 uses the corrected layer set only to bound the weight-scan candidate space.

Relationship to arXiv:2606.04819 (also post-hoc). The paper argues the dominant pool mining binary contains no inference code. §2's T1 result (no labeled pool is model-consistent, ≥99.5% provably not the unmodified stack) points the same way from certificate geometry, and does so independently of the paper's method. But geometry is weaker than the paper's claim: T1-fail proves "not the unmodified published stack", which a modified stack running real inference with different kernels would also fail. So this is directional corroboration of the paper's conclusion, not a proof that pool binaries lack inference code. And it leaves Pearl's published vLLM miner untouched: that is different software (see Keshi's internal protocol notes, The mining stack).

6. Summary (within the pre-registered frame)

• The chain divides sharply: official-pattern traffic is early and almost entirely unattributed; from the MoE fork on it is ≥92% custom.
• No labeled pool is model-consistent, and ≥99.5% of every labeled pool's blocks are provably not the unmodified official stack (T1, era-confound refuted). This is statistics-free, subject only to the 64 unfalsifiable T1-pass labeled blocks.
• The official-pattern traffic that exists is inconsistent with every published model's served-inference mixture (T3), and its m-distribution triggers the pre-registered T4 flags (dominant m > 16,384: "stronger but configuration-dependent"). Per the prereg, this is "consistent with synthetic shape selection or an unpublished model. Not proof of synthetic mining." Superseded in part by the 2026-08-07 erratum in §5: the T3 expected distributions counted layers the published checkpoints quantize FP8 and therefore cannot mine, so the T3 rejections do not support a synthetic-selection reading. T1's results are unaffected.

Limitations (PREREG-001 §9, unchanged)

Attribution is single-wallet per pool. Shape consistency is evidence only against a non-mimicking population: deliberate mimicry defeats it at zero marginal cost; only hash_b weight-commitment verification (Phase 8.3/8.4) resists it. Negative results are coverage-bounded across the four published models. The corpus records shapes, not tensors: nothing here verifies weights. The rank-penalty (E4) stratum is small (155 blocks) and its rank is consensus-constrained (≥128) post-fork, so E4 shares are read cautiously. Independent review (8.6) and the negative control (8.5) precede any registry or API surface.