Verification round — three-lens review of the finalized authorities

EVIDENCE + the applied fix record. Three profile-loaded lenses ran read-only against the corpus at 0fcb4b3d2 after the publication-readiness rewrite. Every finding below marked ✅ was applied in the finalization commit; the authorities are post-fix. This file is the record of what the round found and what changed.

Aggregated verdict

The measurements hold up; the coordination layer needed one more pass — and got it.

  • The reproducibility claim is substantially TRUE, and unusually strong. The debugger lens recomputed ~60 figures by deliberately independent methods (characteristic polynomial + Durand–Kerner instead of the script's power iteration; its own Jacobi PCA instead of trusting any published series). Every substantive number in §6.1–6.4, §10.2, the negative-cell census, and all five grep counts reproduced — most to more decimals than published. Exactly two published figures were off: a mean (0.3842 → 0.3841) and a symmetry count (48/49 → 47/49 cells; one asymmetric pair of 21). Neither moves any conclusion.
  • Two genuine HIGH defects were found in the maths layer, both now fixed:
    1. The reference spectral-radius method was fail-open (F-1). Power iteration from a fixed all-ones start reports gain 0.33 for a two-camp ±1 basis whose true gain is exactly 1.0 — because all-ones is an eigenvector of a smaller eigenvalue. AMMERSE's own figures were never affected (real dominant eigenvalue, non-deficient start), but the method would have shipped into the validator §6.2 calls load-bearing at exactly the boundary case. Fix applied: the committed script now uses seeded random restarts (verified: reports 3.0/2.0 correctly on both counterexamples, AMMERSE unchanged), and §6.2 documents the hazard.
    2. The (1−r) bounding claim confused the spectral norm with the sup-norm (F-2). A per-axis [−1, 1] guarantee is an ℓ∞ statement; the tight two-term constant is 1/(1+r∞) with r∞ = 0.5·max_row_abs_sum/(N−1) — 0.75 for AMMERSE (verified exactly tight: worst admissible ±1 base lands at 1.000000, while (1−r) = 0.8054 overflows to 1.074). And no fixed constant bounds a sum of deltas — the active-set composition must clamp or average first, recorded as an open design consequence. Fix applied in §6.4 + register.
  • The sequence document — the one authority that never had a review pass — turned out to carry the fewest unverifiable claims (all seven of its file:line citations resolve, six exactly; all three precedent anchors reproduce to the file and line), but its coordination layer had five executable blockers, all now fixed: a decision-ID namespace collision (Dn ADR decisions vs D-n design decisions, colliding on different meanings → renamed ADR-Dn); a false "last band entirely behind G2/G3" claim (truth: roughly half — I15 is behind #2467, I10 behind G4, I13/I1c ungated); two missing increments in the critical path (I3c, I8) and an unranked gate (G5); the consistency_check re-point range excluding one function it must re-point (934→917); and two surviving pre-review 49/49 figures.
  • The design doc's §0 ask was priced on the wrong band (fresh-eyes lens): the "Prose-only (§8)" row carried the ranks-1–4 campsite figures, while §8's own costs: field is sequence rank 19 with a 196–228-artefact authoring sweep. Fixed: the row is relabeled, §8 is priced honestly, and the headline "~85% behind two gates" — which reproduced from no column of the effort table — is now "~60–65% of tabulated engineering cost in the final band, roughly half of it behind the two gates, the authoring tail entirely behind #2538".
  • The 1,820-cell figure was never re-derived after §7.4 narrowed the kind set — it counted toolguide/step-contract/glossary_pack artefacts the design excludes. Canonical: ~1,372–1,596 (196–228 × 7). Fixed in both authorities + register.
  • The register's own account of the superseded "~6%" was wrong (it said "undamped three-term", which is 9.65%; the 6% was gain³ = 5.89%, the first-omitted-term magnitude). Fixed — the one row that cannot afford to be wrong is the one explaining a previous error.
  • Corpus-wide figure hygiene: the evidence docs still published superseded numbers with no pointer (the matrix-measurement doc even endorses the three-term composition the authority forbids). Fixed: superseded banners on connascence-matrix-measurement.md, creed-and-values-design-hardened.md, creed-and-values-design-as-proposed.md, manifesto-tier-verdict-and-handover.md, and squad-reports/reviewer.md; six new rows in the authority's §11 register.

Net: nothing found moves the architecture. Everything found moves a number, a label, or a pointer — and the two HIGH maths defects were in the reference implementation and one bounding sentence, not in any adopted result. The published second-order matrices are now committed (_ammerse-second-order.json), closing the one evidence gap where the most consequential empirical claim (the false upstream derivation) was uncheckable.

Fix ledger

# Finding (lens) Severity Status
1 Spectral-radius method fail-open on deficient start vectors (debugger F-1) HIGH ✅ script: seeded random restarts; §6.2 hazard note
2 (1−r) does not bound output in sup-norm; tight constant 0.75; sums unbounded (debugger F-2) HIGH ✅ §6.4 rewritten; register row
3 §0 cost row prices ranks 1–4 as "Prose-only (§8)" (patterns) HIGH ✅ relabeled + §8 priced honestly
4 1,820 cells uses pre-§7.4 superset (patterns) HIGH ✅ ~1,372–1,596 in both authorities + register
5 Dn/D-n namespace collision, 3 pairs mean different things (architect) HIGH ✅ ADR-Dn rename + namespace note
6 "Last band entirely behind G2/G3" false (architect) HIGH ✅ corrected in both authorities
7 "48 of 49 symmetric" arithmetically impossible (architect + debugger F-3) HIGH ✅ "one asymmetric pair of 21 (47/49 cells)" in both docs + JSON
8 Register row 1's "6% = undamped three-term" contradicts §6.1 (patterns + debugger F-6) MEDIUM ✅ corrected attribution (gain³ = 5.89%, first-omitted-term)
9 Second-order matrix not committed — key evidence uncheckable (debugger F-5) MEDIUM ✅ _ammerse-second-order.json committed (raw + normalised)
10 49/49 survivals in sequence doc (architect) MEDIUM ✅ → 42/42 off-diagonal
11 "~85% behind two gates" not derivable (architect + patterns) MEDIUM ✅ → ~60–65% + half-band + tail statement
12 I3c, I8 missing from critical path; G5 unranked (architect) MEDIUM ✅ added; gate-order legend note
13 G3-gated set listed I10 (architect) MEDIUM ✅ I10 → G4
14 consistency_check range 934→917; "7 functions" → 1 dataclass + 5 (architect) MEDIUM ✅
15 JSON missing matrix_version + source_digest (architect) MEDIUM ✅ both added, digest scope stated
16 §6.2 heading "non-degenerate" asserts unproven equivalence (patterns) MEDIUM ✅ heading → "sound for gain < 1"
17 §7.2 "five of seven axes" → six (patterns) MEDIUM ✅
18 §6.5 "32 of 49" → 32 of 42 off-diagonal (patterns) MEDIUM ✅
19 ~47,900 basis unstated; 260-vs-310 divergence unregistered (patterns) MEDIUM ✅ inline basis + register row
20 Stale-figure banners on 5 RECORD/EVIDENCE docs (patterns) MEDIUM ✅
21 PCA convention unstated — correlation-PCA readers get 75.7% (debugger) MEDIUM ✅ convention clause in §10.2
22 Mean 0.3842 → 0.3841 (debugger F-4) LOW ✅ doc + script
23 Script positional cell indexing vs §4's by-id rule (debugger F-7) LOW ✅ axes.index()
24 Reproducibility claim scope (banner bullet 2) (debugger F-5) LOW ✅ scoped per file
25 Sweep figure needs SAMPLES=30000 note (architect) LOW ✅ env-var override + note
26 Walker citation :507 → :507-509; "4 elif" → "4 kind branches"; base-commit bumps; gate-order legend LOW ✅

Deliberately not applied: renumbering gates to path order (stable IDs preferred; legend added instead); moving §9.3/§11 mechanics out of the design doc (the standalone review version handles digestibility separately); the correlation-PSD third constraint (recorded as a candidate in §6.2, not adopted — it is a design choice, not a fix).

What the lenses could not verify (carried forward honestly)

  • Tracker states (#2466, #2467, #2532, #2538, #2591, #2934, the four I2 consumers) — no lens audited them this round; the sequence doc's §4 collision adjudications rest on the planning round's checks.
  • The #2538 rig's liveness — still unverified by anyone; G3 decays with it (G5 exists for exactly this reason).
  • The upstream prose quotations (the "decidedly not minimal" class) — the corpus JSON carries ids and vectors only; the 11-of-12 prose split is verified against numbers, not against the upstream text, which is not committed.
  • The 38→37 tiers of the corpus ladder — upstream-only; the committed JSON starts at 36.
  • The Dirichlet ranking-collapse table in the hardened RECORD doc — no seed or code was ever committed for it; it remains non-reproducible and is banner-flagged as a superseded objection.

Lens verdicts, condensed

  • Architect: "Not as-is, but the gap is one editing pass, not a rewrite… the document that never had a review pass is carrying fewer unverifiable claims than the one that did, because it is the only one that cites line numbers at all." → the editing pass is applied.
  • Patterns (fresh eyes): "Nothing I found moves the architecture, everything I found moves a number… hand it to the human reviewer, with the §0 cost table fixed first." → fixed.
  • Debugger: "Reproducibility of the published figures: verified. Reproducibility of the evidence base, and soundness of the method carried forward into a validator: not yet." → the second-order JSON is committed and the method is restart-hardened; both gaps closed.
Edit this page