Pre-registration, drafted 2026-08-06. Roadmap Phase 8.2 (ROADMAP-certificate-intelligence.md); governed by the Verification standard §4: the hypotheses, tests, thresholds and the interpretation of every possible outcome are fixed here, before the analysis runs at scale. After this document's commit hash is recorded in the dataset manifest, nothing in §§2–7 may change; corrections ship as a PREREG-002 with a stated reason, never as an edit.

Binding language (applies to every downstream artifact). All findings are statements of the form "consistent with" or "inconsistent with" a pre-registered shape signature. Nothing in this analysis can prove that inference occurred, that a model was actually served, or that any computation was fresh. The phrase "was real inference" is prohibited except in this sentence. Per-block accusations are never made; strata are pools and eras, and pools are aggregators of unknown fleets.

Typographic erratum 2026-08-07. This document was edited on 2026-08-07 for typography only: em dashes were rewritten into commas, colons and sentence breaks. No hypothesis, test, threshold, table value, number, or interpretation changed. The freeze is unaffected and stays verifiable rather than merely asserted: the exact bytes this document had when DS-001 was produced are retrievable at the commit its manifest pins, and the diff is inspectable with git diff 11a517a02cedfb97360da95bfaf8e4f114374dfd..d4cb529b8602f1f73bab681f4ae428ba6cfd3fff -- docs/research/PREREG-001-real-inference-detection.md.

Reconciling manifests across this edit. Datasets are append-only, so DS-001's manifest still records preregCommit: 11a517a02ced and that remains the correct pointer to the document its run implemented. Manifests written after 2026-08-07 record d4cb529b8602 instead. The two hashes name the same rules: the only differences are the typography above and this erratum, which is why the pointer was allowed to move at all. A reader reconciling an old dataset with a new one should expect exactly two values here and can confirm their equivalence with the diff command above. Any future change to §§2-7 ships as PREREG-002-style new document, never as a further pointer move.

1. Question and hypotheses

Pearl's official miner is a vLLM plugin: real inference where quantized linear layers double as proof-of-work, recording the true GEMM dimensions in the certificate (pearl-notes.md §The mining stack). Synthetic miners choose dimensions freely. The corpus therefore admits shape-level discrimination.

• H1: some portion of the chain's certificates is consistent with the unmodified official stack mining published pearl-ai models.
• H2: some portion is provably not the unmodified official stack (T1-fail is a hard falsifier; the reverse direction proves nothing).
• H3: within official-pattern traffic, per-stratum (k, n) frequencies match the MAC-share mixture a single served model predicts (the roadmap's T3): "a distributional fingerprint a synthetic miner cannot cheaply fake, because faking it means actually doing the layer mix").

2. Population, eras, strata

• Population: canonical blocks with a decoded cert_public_params row at decoder version 1. Orphaned blocks' rows are excluded (pre-registered; they remain in the corpus for other work). Height 0 is class EMPTY and is excluded from all denominators.
• Eras (activation-at-height semantics, from chaincfg + registry/PCCR.md): E1 pre-moe 1–71,934 · E2 moe-window 71,935–91,629 · E3 dense-only 91,630–96,250 · E4 rank-penalty 96,251–tip-at-snapshot. The v1.2.1 IsMoE semantic fix has no height and no effect here (MoE-ness is moe_e > 0 on decoded params; inconsistent tails already fail the Phase 6 parse). Recorded as a caveat, not a boundary.
• Strata: pool × era. Pool from the coinbase address against pool_labels; the unattributed remainder is its own stratum, always shown, never redistributed.

3. T1: official-stack byte test (per block, binary)

T1 passes iff ALL of:

1. rank ∈ {64, 128}, the only noise ranks compiled in the published kernels (pearl-gemm-build-utils/.../default_compiled_kernels.py).
2. rows_pattern == 0x070100000000 byte-exact, decoding to (stride 8, len 2),(16,1),(16,1): two rows 8 apart, one sm90 wgmma accumulator fragment.
3. cols_pattern == 0x0001031f0000 byte-exact, decoding to (1,2),(8,32),(256,1): 32 stride-8 pairs = 64 columns.
4. m ≥ 1024 ∧ n ≥ 256 ∧ k ≥ 1024, the vLLM miner's engagement floor (vllm-miner/config.yaml).

Span matching is prohibited. The fork-census window alone carries five distinct byte encodings with h·w = 128, including a contiguous 2×64 pattern (rows 0x000100000000, cols 0x003f00000000; e.g. height 96,334) produced by a different kernel; only the exact bytes identify the published fragment.

Era note: in E4 (rank-penalty), rank 64 is consensus-invalid, so E4 T1 blocks necessarily carry rank 128. That is expected, not an anomaly.

Asymmetric semantics (binding): because MINER_* environment variables can override rank, patterns and tiles, T1-fail proves the block was not produced by the unmodified published stack; T1-pass proves nothing.

4. T2: layer-geometry test (per block)

A block matches layer row (model, layer, tp) iff its (n, k) equals the row's pair after tensor-parallel division. Column-parallel layers (qkv_fused, gate_up_fused) divide n by tp; row-parallel layers (o, down) divide k by tp, with tp ∈ {1, 2, 4, 8} and the divided pair still satisfying n ≥ 256 ∧ k ≥ 1024 (rows pruned by the floor at a given tp simply do not enter the grid at that tp). Dense rows additionally require moe_e = 0; the Qwen3 expert row requires (moe_e, moe_top_k) = (128, 8). T2 passes iff the match set is non-empty. Ambiguity is recorded, never resolved: every match (with its provenance tag) is emitted per block; a multi-match block counts once as "any-model consistent".

Frozen layer tables (tp = 1; n = out-features, k = in-features, per gemm_operators.py)

Provenance tags: anchored (dimension pair observed dominating the chain and/or derivable from verified config), documented (stated in pearl-notes.md F4), derived (computed from documented architecture, not independently verified).

† Gemma attention dims verified live from the published HF config on 2026-08-06 (text_config: heads 32 × head_dim 256 → attn 8192; kv heads 16 → 2×4096; qkv_fused n = 8192 + 8192 = 16384). An earlier dossier draft listed o as (5376, 5376), which is wrong, since Gemma-4's attention dimension (8192) differs from its hidden size (5376). Whether Gemma's attention layers are in the INT7 group is assumed by analogy with the Llama models and NOT independently verified. These two rows affect T2 class shares only (Gemma is T3-excluded) and are filterable by tag.

Deliberate exclusions, with reasons: Qwen3 expert down (2048, 768), an FP8 block-scaled path, not INT7-mined, and k = 768 < 1024 regardless; Qwen3 attention layers, INT7 membership unverified (a non-classifying sensitivity table reports qkv (5120, 2048) and o (2048, 4096) counts separately). Fused pairs are used for the Llama models because vLLM's QKVParallelLinear / MergedColumnParallelLinear execute fused GEMMs. The dossier's unfused F4 listing is corrected by errata in the same commit as this document.

5. Classification (per block)

plus EMPTY (genesis only). Reported per stratum as counts and shares.

6. T3: layer-mixture test (per stratum × hypothesis)

• Decision statistic: total-variation distance, an effect size with pre-registrable thresholds. Chi-square/G reject any point null at corpus sample sizes on trivial deviations; the G-statistic G = 2·Σ O_c·ln(O_c/E_c) is computed and reported descriptively only.
• Hypothesis grid, fixed: H = {Llama-3.1-8B, Llama-3.3-70B} × tp {1, 2, 4, 8}, plus Qwen3-30B-A3B × tp {1, 2, 4} (its expert row prunes at tp = 8: n = 192 < 256). Gemma is T3-excluded (its attention rows' INT7 membership is assumed, not verified → no defensible mixture). Every grid cell is reported for every stratum: no best-fit selection, no post-hoc hypotheses. That is the forking-paths control.
• Per-cell support and pruning: a grid cell's category support is the model's layer rows that survive the min-dims floor at that tp, and the expected vector q is renormalized over the surviving support. Concretely the only Llama pruning is (8B, tp = 8), where o drops (k = 512 < 1024) and q renormalizes to gate_up 0.5833 / down 0.2917 / qkv 0.1250.
• Base population B_s = T1-passing blocks of stratum s; N = |B_s|. N < 200 → verdict INSUFFICIENT_SAMPLE.
• Primary form: categories = support(H) ∪ {OTHER}; observed p̂(c) = count(c)/N; expected q(c) = MAC shares with q(OTHER) = 0; TV = ½·Σ|p̂ − q|.
• Conditional form (mixed-fleet view): restrict to support(H), renormalize; requires matched count ≥ 200. Both forms always reported.
• Expected MAC shares, frozen arithmetic: per-token MACs of layer l are n_l · k_l; every mined layer executes once per token per shard, so token counts and tp cancel: share_l = n_l·k_l / Σ n·k (tp-invariant except where the min-dims floor prunes a layer at a given tp, in which case q renormalizes over the survivors as defined above). If block yield were ∝ m·n·k_eff instead of per-token MACs, the prediction is unchanged for T1 blocks: k_eff = k exactly at rank ≤ 128 for every table layer (all k are multiples of 128, the same k_eff the v1.3.0 bound multiplier (k_eff/rank)·128 uses, pearl-notes §rank-penalty rule).

Qwen3 semantics, fixed: with a single category the primary form is well defined and applies normally, with TV_primary = 1 − p̂(expert pair), so TV ≤ 0.10 means ≥90% of the stratum's T1 blocks carry (1536, 2048) with the (128, 8) trailer, which CAN trigger outcome row 1. The conditional form is degenerate (always 0) and is reported as DEGENERATE, never as a verdict.

• Thresholds (frozen): TV ≤ 0.10 → MIXTURE_CONSISTENT; 0.10 < TV ≤ 0.25 → INDETERMINATE; TV > 0.25 → MIXTURE_INCONSISTENT. Justification: under a true mixture, E|p̂−p| ≈ √(2p(1−p)/(πN)) per category, giving E[TV] ≈ 0.043 at N = 200 for the Llama share vectors, so the 0.10 line sits ≈2.3× above null expectation (≈0.019 at N = 1000).

Outcome → interpretation table (copied verbatim into OBS-004)

Quantifier rule: DEGENERATE and INSUFFICIENT_SAMPLE values are excluded from every "for some H" / "for all H" quantification below.

7. T4: m-variability (per stratum; base = T1∧T2 blocks; N ≥ 50)

Report per stratum: distinct-m share |{m}|/N; mode-m share; power-of-two-m share; share with m > 16,384. Flags: CONSTANT_M iff mode-m share ≥ 0.90; POW2_M iff pow2 share ≥ 0.90.

Pre-written caveat (binding): vLLM chunked prefill yields constant power-of-two m at configuration values (2048/4096/8192/16384), so CONSTANT_M at a plausible chunk size is weak evidence of anything. Constant m at non-chunk values, or dominant m > 16,384, is stronger but still configuration-dependent (max_num_batched_tokens is operator-chosen). MoE routing thins per-expert m in ways this analysis does not model, a documented unknown.

8. Procedure, dataset, and freeze mechanics

1. This document commits first; its git hash is the freeze reference.
2. internal/certclass implements §§3–7 as pure functions; its constants cite this document's sections, and a unit test asserts the computed MAC shares equal the frozen literals above, including the renormalized (8B, tp = 8) vector 0.5833/0.2917/0.1250 (table drift = test failure).
3. keshictl classify refuses to run unless the corpus is complete (certParams missing = 0) and walks one REPEATABLE READ snapshot, emitting DS-001 (docs/research/datasets/DS-001-classification-v1/): blocks.jsonl.gz (per block: height, hash, pool, era, class, per-test results, matches with provenance tags), summary.json (per-stratum class counts, the full T3 grid, T4 stats; thresholds and versions echoed), MANIFEST.md (sha256s, snapshot tip height+hash, tool + decoder + prereg commit). Byte-determinism is a tested property.
4. DS-001 is append-only per roadmap §6.5: re-runs are -v2 with an errata note, never edits.
5. OBS-004 fills §6's interpretation table with measured numbers. Any observation not derivable from the pre-registered tables goes in a clearly labeled "Exploratory (post-hoc)" section. Adversarial review precedes publication (roadmap 8.6); negative-control obligations (8.5) apply before any registry or API surface ships.

9. Known limitations (stated in advance)

• Attribution is single-known-wallet per pool; proxy payouts misattribute.
• The layer tables cover four published models; an unpublished or fine-tuned model with different dims lands in OFFICIAL_PATTERN_NONMODEL/CUSTOM by construction. Negative results are coverage-bounded, never proof of absence.
• T1's engagement floor and T4's chunk sizes come from the default config; operators can change both.
• Gemma rows are partially derived; they affect T2 class shares only and are filterable in re-analysis via provenance tags.
• The corpus records shapes, not tensors: nothing here verifies weights (that is Phase 8.3/8.4, hash_b recomputation).
• Deliberate mimicry defeats this analysis at zero marginal cost. Under the yield model above, a synthetic miner drawing table shapes in MAC proportions with the official bytes, rank 128 and power-of-two m lands OFFICIAL_CONSISTENT and MIXTURE_CONSISTENT while sacrificing nothing. Shape consistency is therefore evidence only against the current, non-adversarial population; only weight-commitment verification (hash_b, Phase 8.3/8.4) resists mimicry.