<?xml version="1.0" encoding="utf-8"?>
<entry xmlns="http://www.w3.org/2005/Atom" xml:base="https://docs.spec-kitty.ai/plans/doctrine/squad-reports/review-round-2026-07-26/">
  <id>https://docs.spec-kitty.ai/plans/doctrine/squad-reports/review-round-2026-07-26/</id>
  <title>Verification round — three-lens review of the finalized authorities</title>
  <updated>2026-07-26T00:00:00.000Z</updated>
  <author><name>Spec Kitty</name></author>
  <link rel="alternate" type="text/html" href="https://docs.spec-kitty.ai/plans/doctrine/squad-reports/review-round-2026-07-26/"/>
  <summary>Final review squad before handover: sequence executability, fresh-eyes fix verification, and adversarial number-by-number reproduction. Verdict plus the applied fix list.</summary>
  <content type="html">&lt;span id=&quot;verification-round--three-lens-review-of-the-finalized-authorities&quot;&gt;&lt;/span&gt;
&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;Retired (deprecated).&lt;/strong&gt; Superseded by the canonical creed AUTHORITY doc &lt;a href=&quot;../foundational-values-and-creed.md&quot;&gt;foundational-values-and-creed.md&lt;/a&gt;. Preserved as a historical record.&lt;/p&gt;
&lt;/blockquote&gt;
&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;EVIDENCE + the applied fix record.&lt;/strong&gt; Three profile-loaded lenses ran read-only against the
corpus at &lt;code dir=&quot;auto&quot;&gt;0fcb4b3d2&lt;/code&gt; after the publication-readiness rewrite. Every finding below marked ✅ was
&lt;strong&gt;applied in the finalization commit&lt;/strong&gt;; the authorities are post-fix. This file is the record of
what the round found and what changed.&lt;/p&gt;
&lt;/blockquote&gt;
&lt;h2 id=&quot;aggregated-verdict&quot;&gt;Aggregated verdict&lt;/h2&gt;
&lt;p&gt;&lt;strong&gt;The measurements hold up; the coordination layer needed one more pass — and got it.&lt;/strong&gt;&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;The reproducibility claim is substantially TRUE, and unusually strong.&lt;/strong&gt; 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). &lt;strong&gt;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.&lt;/strong&gt; Exactly two
published figures were off: a mean (0.3842 → &lt;strong&gt;0.3841&lt;/strong&gt;) and a symmetry count (48/49 → &lt;strong&gt;47/49
cells; one asymmetric pair of 21&lt;/strong&gt;). Neither moves any conclusion.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Two genuine HIGH defects were found in the maths layer, both now fixed:&lt;/strong&gt;
&lt;ol&gt;
&lt;li&gt;&lt;strong&gt;The reference spectral-radius method was fail-open&lt;/strong&gt; (F-1). Power iteration from a fixed
all-ones start reports gain &lt;strong&gt;0.33&lt;/strong&gt; for a two-camp ±1 basis whose true gain is exactly
&lt;strong&gt;1.0&lt;/strong&gt; — because all-ones is an eigenvector of a &lt;em&gt;smaller&lt;/em&gt; 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. &lt;strong&gt;Fix
applied:&lt;/strong&gt; 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.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;The &lt;code dir=&quot;auto&quot;&gt;(1−r)&lt;/code&gt; bounding claim confused the spectral norm with the sup-norm&lt;/strong&gt; (F-2). A per-axis
[−1, 1] guarantee is an ℓ∞ statement; the tight two-term constant is &lt;code dir=&quot;auto&quot;&gt;1/(1+r∞)&lt;/code&gt; with
&lt;code dir=&quot;auto&quot;&gt;r∞ = 0.5·max_row_abs_sum/(N−1)&lt;/code&gt; — &lt;strong&gt;0.75&lt;/strong&gt; for AMMERSE (verified exactly tight: worst
admissible ±1 base lands at 1.000000, while &lt;code dir=&quot;auto&quot;&gt;(1−r) = 0.8054&lt;/code&gt; overflows to &lt;strong&gt;1.074&lt;/strong&gt;). And no
fixed constant bounds a &lt;em&gt;sum&lt;/em&gt; of deltas — the active-set composition must clamp or average
first, recorded as an open design consequence. &lt;strong&gt;Fix applied&lt;/strong&gt; in §6.4 + register.&lt;/li&gt;
&lt;/ol&gt;
&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;The sequence document — the one authority that never had a review pass — turned out to carry
the fewest unverifiable claims&lt;/strong&gt; (all seven of its &lt;code dir=&quot;auto&quot;&gt;file:line&lt;/code&gt; citations resolve, six exactly;
all three precedent anchors reproduce to the file and line), &lt;strong&gt;but its coordination layer had
five executable blockers&lt;/strong&gt;, all now fixed: a decision-ID namespace collision (&lt;code dir=&quot;auto&quot;&gt;Dn&lt;/code&gt; ADR decisions
vs &lt;code dir=&quot;auto&quot;&gt;D-n&lt;/code&gt; design decisions, colliding on different meanings → renamed &lt;strong&gt;&lt;code dir=&quot;auto&quot;&gt;ADR-Dn&lt;/code&gt;&lt;/strong&gt;); a false
“last band entirely behind G2/G3” claim (truth: roughly half — I15 is behind &lt;code dir=&quot;auto&quot;&gt;#2467&lt;/code&gt;, I10 behind
G4, I13/I1c ungated); two missing increments in the critical path (I3c, I8) and an unranked gate
(G5); the &lt;code dir=&quot;auto&quot;&gt;consistency_check&lt;/code&gt; re-point range excluding one function it must re-point
(934→&lt;strong&gt;917&lt;/strong&gt;); and two surviving pre-review &lt;code dir=&quot;auto&quot;&gt;49/49&lt;/code&gt; figures.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;The design doc’s §0 ask was priced on the wrong band&lt;/strong&gt; (fresh-eyes lens): the “Prose-only (§8)”
row carried the ranks-1–4 campsite figures, while §8’s own &lt;code dir=&quot;auto&quot;&gt;costs:&lt;/code&gt; field is sequence rank 19
with a 196–228-artefact authoring sweep. &lt;strong&gt;Fixed:&lt;/strong&gt; 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 “&lt;strong&gt;~60–65% of tabulated engineering cost in the final band, roughly half of it behind the
two gates, the authoring tail entirely behind &lt;code dir=&quot;auto&quot;&gt;#2538&lt;/code&gt;&lt;/strong&gt;”.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;The 1,820-cell figure was never re-derived after §7.4 narrowed the kind set&lt;/strong&gt; — it counted
toolguide/step-contract/glossary_pack artefacts the design excludes. Canonical: &lt;strong&gt;~1,372–1,596&lt;/strong&gt;
(196–228 × 7). Fixed in both authorities + register.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;The register’s own account of the superseded “~6%” was wrong&lt;/strong&gt; (it said “undamped three-term”,
which is 9.65%; the 6% was &lt;code dir=&quot;auto&quot;&gt;gain³ = 5.89%&lt;/code&gt;, the first-omitted-term magnitude). Fixed — the one
row that cannot afford to be wrong is the one explaining a previous error.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Corpus-wide figure hygiene:&lt;/strong&gt; the evidence docs still published superseded numbers with no
pointer (the matrix-measurement doc even &lt;em&gt;endorses the three-term composition the authority
forbids&lt;/em&gt;). &lt;strong&gt;Fixed:&lt;/strong&gt; superseded banners on &lt;code dir=&quot;auto&quot;&gt;connascence-matrix-measurement.md&lt;/code&gt;,
&lt;code dir=&quot;auto&quot;&gt;creed-and-values-design-hardened.md&lt;/code&gt;, &lt;code dir=&quot;auto&quot;&gt;creed-and-values-design-as-proposed.md&lt;/code&gt;,
&lt;code dir=&quot;auto&quot;&gt;manifesto-tier-verdict-and-handover.md&lt;/code&gt;, and &lt;code dir=&quot;auto&quot;&gt;squad-reports/reviewer.md&lt;/code&gt;; six new rows in the
authority’s §11 register.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;&lt;strong&gt;Net:&lt;/strong&gt; nothing found moves the architecture. Everything found moves a number, a label, or a
pointer — and the two HIGH maths defects were in the &lt;em&gt;reference implementation and one bounding
sentence&lt;/em&gt;, not in any adopted result. The published second-order matrices are now committed
(&lt;code dir=&quot;auto&quot;&gt;_ammerse-second-order.json&lt;/code&gt;), closing the one evidence gap where the most consequential
empirical claim (the false upstream derivation) was uncheckable.&lt;/p&gt;
&lt;h2 id=&quot;fix-ledger&quot;&gt;Fix ledger&lt;/h2&gt;







































































































































































&lt;table&gt;&lt;thead&gt;&lt;tr&gt;&lt;th&gt;#&lt;/th&gt;&lt;th&gt;Finding (lens)&lt;/th&gt;&lt;th&gt;Severity&lt;/th&gt;&lt;th&gt;Status&lt;/th&gt;&lt;/tr&gt;&lt;/thead&gt;&lt;tbody&gt;&lt;tr&gt;&lt;td&gt;1&lt;/td&gt;&lt;td&gt;Spectral-radius method fail-open on deficient start vectors (debugger F-1)&lt;/td&gt;&lt;td&gt;HIGH&lt;/td&gt;&lt;td&gt;✅ script: seeded random restarts; §6.2 hazard note&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;2&lt;/td&gt;&lt;td&gt;&lt;code dir=&quot;auto&quot;&gt;(1−r)&lt;/code&gt; does not bound output in sup-norm; tight constant 0.75; sums unbounded (debugger F-2)&lt;/td&gt;&lt;td&gt;HIGH&lt;/td&gt;&lt;td&gt;✅ §6.4 rewritten; register row&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;3&lt;/td&gt;&lt;td&gt;§0 cost row prices ranks 1–4 as “Prose-only (§8)” (patterns)&lt;/td&gt;&lt;td&gt;HIGH&lt;/td&gt;&lt;td&gt;✅ relabeled + §8 priced honestly&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;4&lt;/td&gt;&lt;td&gt;1,820 cells uses pre-§7.4 superset (patterns)&lt;/td&gt;&lt;td&gt;HIGH&lt;/td&gt;&lt;td&gt;✅ ~1,372–1,596 in both authorities + register&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;5&lt;/td&gt;&lt;td&gt;&lt;code dir=&quot;auto&quot;&gt;Dn&lt;/code&gt;/&lt;code dir=&quot;auto&quot;&gt;D-n&lt;/code&gt; namespace collision, 3 pairs mean different things (architect)&lt;/td&gt;&lt;td&gt;HIGH&lt;/td&gt;&lt;td&gt;✅ &lt;code dir=&quot;auto&quot;&gt;ADR-Dn&lt;/code&gt; rename + namespace note&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;6&lt;/td&gt;&lt;td&gt;“Last band entirely behind G2/G3” false (architect)&lt;/td&gt;&lt;td&gt;HIGH&lt;/td&gt;&lt;td&gt;✅ corrected in both authorities&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;7&lt;/td&gt;&lt;td&gt;“48 of 49 symmetric” arithmetically impossible (architect + debugger F-3)&lt;/td&gt;&lt;td&gt;HIGH&lt;/td&gt;&lt;td&gt;✅ “one asymmetric pair of 21 (47/49 cells)” in both docs + JSON&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;8&lt;/td&gt;&lt;td&gt;Register row 1’s “6% = undamped three-term” contradicts §6.1 (patterns + debugger F-6)&lt;/td&gt;&lt;td&gt;MEDIUM&lt;/td&gt;&lt;td&gt;✅ corrected attribution (&lt;code dir=&quot;auto&quot;&gt;gain³ = 5.89%&lt;/code&gt;, first-omitted-term)&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;9&lt;/td&gt;&lt;td&gt;Second-order matrix not committed — key evidence uncheckable (debugger F-5)&lt;/td&gt;&lt;td&gt;MEDIUM&lt;/td&gt;&lt;td&gt;✅ &lt;code dir=&quot;auto&quot;&gt;_ammerse-second-order.json&lt;/code&gt; committed (raw + normalised)&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;10&lt;/td&gt;&lt;td&gt;&lt;code dir=&quot;auto&quot;&gt;49/49&lt;/code&gt; survivals in sequence doc (architect)&lt;/td&gt;&lt;td&gt;MEDIUM&lt;/td&gt;&lt;td&gt;✅ → 42/42 off-diagonal&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;11&lt;/td&gt;&lt;td&gt;“~85% behind two gates” not derivable (architect + patterns)&lt;/td&gt;&lt;td&gt;MEDIUM&lt;/td&gt;&lt;td&gt;✅ → ~60–65% + half-band + tail statement&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;12&lt;/td&gt;&lt;td&gt;I3c, I8 missing from critical path; G5 unranked (architect)&lt;/td&gt;&lt;td&gt;MEDIUM&lt;/td&gt;&lt;td&gt;✅ added; gate-order legend note&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;13&lt;/td&gt;&lt;td&gt;G3-gated set listed I10 (architect)&lt;/td&gt;&lt;td&gt;MEDIUM&lt;/td&gt;&lt;td&gt;✅ I10 → G4&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;14&lt;/td&gt;&lt;td&gt;&lt;code dir=&quot;auto&quot;&gt;consistency_check&lt;/code&gt; range 934→917; “7 functions” → 1 dataclass + 5 (architect)&lt;/td&gt;&lt;td&gt;MEDIUM&lt;/td&gt;&lt;td&gt;✅&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;15&lt;/td&gt;&lt;td&gt;JSON missing &lt;code dir=&quot;auto&quot;&gt;matrix_version&lt;/code&gt; + &lt;code dir=&quot;auto&quot;&gt;source_digest&lt;/code&gt; (architect)&lt;/td&gt;&lt;td&gt;MEDIUM&lt;/td&gt;&lt;td&gt;✅ both added, digest scope stated&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;16&lt;/td&gt;&lt;td&gt;§6.2 heading “non-degenerate” asserts unproven equivalence (patterns)&lt;/td&gt;&lt;td&gt;MEDIUM&lt;/td&gt;&lt;td&gt;✅ heading → “sound for gain &amp;#x3C; 1”&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;17&lt;/td&gt;&lt;td&gt;§7.2 “five of seven axes” → six (patterns)&lt;/td&gt;&lt;td&gt;MEDIUM&lt;/td&gt;&lt;td&gt;✅&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;18&lt;/td&gt;&lt;td&gt;§6.5 “32 of 49” → 32 of 42 off-diagonal (patterns)&lt;/td&gt;&lt;td&gt;MEDIUM&lt;/td&gt;&lt;td&gt;✅&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;19&lt;/td&gt;&lt;td&gt;~47,900 basis unstated; 260-vs-310 divergence unregistered (patterns)&lt;/td&gt;&lt;td&gt;MEDIUM&lt;/td&gt;&lt;td&gt;✅ inline basis + register row&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;20&lt;/td&gt;&lt;td&gt;Stale-figure banners on 5 RECORD/EVIDENCE docs (patterns)&lt;/td&gt;&lt;td&gt;MEDIUM&lt;/td&gt;&lt;td&gt;✅&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;21&lt;/td&gt;&lt;td&gt;PCA convention unstated — correlation-PCA readers get 75.7% (debugger)&lt;/td&gt;&lt;td&gt;MEDIUM&lt;/td&gt;&lt;td&gt;✅ convention clause in §10.2&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;22&lt;/td&gt;&lt;td&gt;Mean 0.3842 → 0.3841 (debugger F-4)&lt;/td&gt;&lt;td&gt;LOW&lt;/td&gt;&lt;td&gt;✅ doc + script&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;23&lt;/td&gt;&lt;td&gt;Script positional cell indexing vs §4’s by-id rule (debugger F-7)&lt;/td&gt;&lt;td&gt;LOW&lt;/td&gt;&lt;td&gt;✅ &lt;code dir=&quot;auto&quot;&gt;axes.index()&lt;/code&gt;&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;24&lt;/td&gt;&lt;td&gt;Reproducibility claim scope (banner bullet 2) (debugger F-5)&lt;/td&gt;&lt;td&gt;LOW&lt;/td&gt;&lt;td&gt;✅ scoped per file&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;25&lt;/td&gt;&lt;td&gt;Sweep figure needs &lt;code dir=&quot;auto&quot;&gt;SAMPLES=30000&lt;/code&gt; note (architect)&lt;/td&gt;&lt;td&gt;LOW&lt;/td&gt;&lt;td&gt;✅ env-var override + note&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;26&lt;/td&gt;&lt;td&gt;Walker citation &lt;code dir=&quot;auto&quot;&gt;:507&lt;/code&gt; → &lt;code dir=&quot;auto&quot;&gt;:507-509&lt;/code&gt;; “4 elif” → “4 kind branches”; base-commit bumps; gate-order legend&lt;/td&gt;&lt;td&gt;LOW&lt;/td&gt;&lt;td&gt;✅&lt;/td&gt;&lt;/tr&gt;&lt;/tbody&gt;&lt;/table&gt;
&lt;p&gt;&lt;strong&gt;Deliberately not applied:&lt;/strong&gt; 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).&lt;/p&gt;
&lt;h2 id=&quot;what-the-lenses-could-not-verify-carried-forward-honestly&quot;&gt;What the lenses could not verify (carried forward honestly)&lt;/h2&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Tracker states&lt;/strong&gt; (&lt;code dir=&quot;auto&quot;&gt;#2466&lt;/code&gt;, &lt;code dir=&quot;auto&quot;&gt;#2467&lt;/code&gt;, &lt;code dir=&quot;auto&quot;&gt;#2532&lt;/code&gt;, &lt;code dir=&quot;auto&quot;&gt;#2538&lt;/code&gt;, &lt;code dir=&quot;auto&quot;&gt;#2591&lt;/code&gt;, &lt;code dir=&quot;auto&quot;&gt;#2934&lt;/code&gt;, the four I2 consumers)
— no lens audited them this round; the sequence doc’s §4 collision adjudications rest on the
planning round’s checks.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;The &lt;code dir=&quot;auto&quot;&gt;#2538&lt;/code&gt; rig’s liveness&lt;/strong&gt; — still unverified by anyone; G3 decays with it (G5 exists for
exactly this reason).&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;The upstream prose quotations&lt;/strong&gt; (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.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;The 38→37 tiers of the corpus ladder&lt;/strong&gt; — upstream-only; the committed JSON starts at 36.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;The Dirichlet ranking-collapse table&lt;/strong&gt; 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.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id=&quot;lens-verdicts-condensed&quot;&gt;Lens verdicts, condensed&lt;/h2&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Architect:&lt;/strong&gt; &lt;em&gt;“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.”&lt;/em&gt; → the editing pass is applied.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Patterns (fresh eyes):&lt;/strong&gt; &lt;em&gt;“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.”&lt;/em&gt; → fixed.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Debugger:&lt;/strong&gt; &lt;em&gt;“Reproducibility of the published figures: verified. Reproducibility of the
evidence base, and soundness of the method carried forward into a validator: not yet.”&lt;/em&gt; → the
second-order JSON is committed and the method is restart-hardened; both gaps closed.&lt;/li&gt;
&lt;/ul&gt;</content>
</entry>
