Skip to content

Research · OBS-003

OBS-003: Phase 7 acceptance at the fork boundary, and first full-corpus census

The acceptance test for Keshi’s certificate-arithmetic series: the new accounting had to reproduce the fork discontinuity OBS-001 measured, or it would not ship. It passed, and the same census refuted the assumption that Pearl’s MoE window went unused.

RecordedRecorded 2026-08-06docs/research/OBS-003-phase7-acceptance-and-corpus-census.md (external link)@7f9d428c396eRendered from the source repository at the commit shown. The content is committed alongside this site, never fetched live.

Dated observation note. Recorded 2026-08-06 (tip ~96,395), from the deployed Phase 7 API (/v1/metrics/{series}, /v1/certs/census) over the complete decoded corpus (96,397 canonical certificates, decoder v1).

1. The fork-discontinuity acceptance (roadmap Phase 7, binding): PASSED

Prediction: the new arithmetic accounting must show a clean discontinuity at the rank-penalty activation (height 96,251) of the magnitude OBS-001 measured, or it must not ship. Boundary-local (last/first 100 blocks around 96,251, 7d window):

Series Pre Post Verdict
rank (mode) 1024 in 49/100 128 in 83/100 PASS: reproduces OBS-001 (49% / 84%)
declared-arithmetic (median) 4.40e13 MACs 4.33e12 MACs PASS: 10× collapse
proven-arithmetic (median) 2.10e6 MACs 5.24e5 MACs PASS: 4× clean step (1.05e6 → 5.24e5 across 25-block medians)
attestation-ratio (median, per block) 9.54e-8 1.21e-7 PASS: steps up (m·n shrank)

The 7d MAC-weighted aggregate ratio is 5.98e-10, far below the per-block median because Σdeclared is dominated by the largest declared shapes. Exactly the median-vs-MAC-weighted distinction metrics.md#attestation-ratio requires quoting explicitly.

2. The MoE window WAS used: hypothesis refuted

moeBlocks = 1,929 over the entire chain. MoE certificates are only consensus-valid inside the window 71,935–91,630 (19,695 blocks), so ~9.8% of that window carried a production MoE proof. This refutes the working hypothesis (roadmap Open questions; echoed in OBS-001) that the MoE capability went essentially unused. The governance story is therefore sharper, not weaker: a hard-forked capability with real production usage was retracted ~44 days later by a softfork that no PIP documents (PCCR-0003).

Errata applied to OBS-001, whose zero-MoE observation was over a post-window sample where zero is consensus-required and supports no usage inference.

3. The chain's most common shape matches the official stack AND a published model

Top declared shape over the whole corpus:

m 32768 · n 57344 · k 8192 · rank 128 · tile 2×64: 15,902 blocks (~16.5%)
  • tile 2×64 is the official kernel's exact wgmma fragment (h=2, w=64; the F9 fingerprint), and rank 128 is a compiled official rank;
  • (n 57344, k 8192) is Llama-3.3-70B-Instruct-pearl's fused gate/up layer (pearl-notes §mining stack, F4).

So roughly a sixth of the chain is consistent with official-stack mining of a published model. That is the strongest real-inference signal yet, and the n=3 fixture sample (which matched no model) badly under-represented the corpus. Per the roadmap's binding reframing: this is "consistent with the published vLLM-plugin shape signature", not a claim that inference occurred; the discrimination bound is Phase 8's T3 mixture test. Caveat: m = 32,768 exceeds typical vLLM batch bounds (~16k), a Phase 8 question, not a conclusion.

Also new against the fixture-era picture:

  • Full-chain rank histogram: 64 → 28,417 · 128 → 46,611 · 256 → 10,863 · 512 → 5,836 · 1024 → 4,668 · 32 → 1 · 0 → 1 (genesis). The early chain mined heavily at rank 64 (an official compiled kernel), and exactly one block used rank 32, the old reference default.
  • Per-pool rank crosstab is live for five labeled pools + the unattributed remainder (always shown, never redistributed).

Method / caveats

  • All figures from the public API of the deployed release ba3db02ecb3b-w81398ae; reproducible with two curl commands per table.
  • Census is whole-window aggregate (REPEATABLE READ snapshot); series carry ≤5,000 newest points per window; boundary-local slices used above.
  • Shape-consistency statements are geometry only (T1/T2 screens); no distributional test has run yet (Phase 8.2), and no claim of served inference is made or implied.