FINAL AUDIT REPORT
Protein interaction-response release v1.0.5
16 September 2026
Physical Sufficiency and Predictive Identifiability in Amino-Acid
Transfer-Energy Representations

SCOPE AND CENTRAL RESULT
This scientific revision separates finite physical sufficiency from
unregularized affine predictive identifiability under a specified validation
protocol. It is a retrospective, self-contained methodological study.
Local execution plans record analysis order; no independent preregistration
before original response exposure is asserted.

NEW COMPUTATIONAL CERTIFICATES
1. Exact rational deletion enumeration of A=[1,q2]: rank 6, all 20 singleton
   deletions pass, and only Cys/Met fails among all 190 unordered pairs
   (remaining rank 5). Thus cospark(A)=2: ordinary LOO safe, nested LOO unsafe.
   This failure persists if the four-coordinate predictive capacity limit
   is relaxed. Exactness refers to the recorded finite decimal ledger.
2. Two complete synchronized H2 null campaigns: 10,000 draws each, seed
   2026091601, one common residue permutation for every endpoint, complete
   nested shared selection over the same 47 safe classes. Observed aggregate
   Q2: 0.5490920693730571 and 0.5756109186180024. Both have zero exceedances,
   plus-one p=1/10001 and Holm-adjusted p=2/10001 across the two new tests.
   The two-sided 95% binomial interval for each unknown exceedance probability
   is [0,0.000368819914619]. Zero events do not imply a zero p-value.
   Two permutations per campaign were independently checked with explicit
   inner/outer refits. Endpoint covariance is preserved. The null concerns
   exchangeable whole-profile alignment, not chemistry-stratified
   exchangeability, superiority of shared over separate models, unique
   geometry or correction for prior manual pipeline discovery.
3. External structure-only FreeSolv audit: all 642 structures and a locally
   specified subset of 40 numeric identifiers. Both atom-count designs have
   rank 9 and pass every ordinary LOO deletion. Each of seven overlapping
   element-family holdouts loses rank; every corresponding held-out row is
   unsupported. Invertible basis changes alter minimum-norm prediction
   operators. Intercept-plus-carbon controls preserve rank and operator
   invariance. No response values are used. This is a deliberately diagnostic
   external demonstration, not a prevalence or predictive-benefit benchmark.

FULL SCIENTIFIC EXECUTION
reproduce_everything.py completed successfully on 16 September 2026.
verification/COMPLETE_SCIENTIFIC_REPLAY.log contains the complete transcript:
- exact structural class counts 415/262/131, 103/93/47, 207/161/89;
- matched nested Ridge/PCR/PLS baselines and deterministic endpoint replay;
- four 50,000-draw H1 pipeline nulls and the separate fixed-class stress null;
- 20,000 WW uncertainty perturbations;
- 20,000 independent H2 alignment draws and 9,000 local sensitivity draws;
- exact quantitative certificates and singleton-Arg basis counterexample;
- historical Cys coding contrast and five million fixed-basis permutations;
- all new q2, synchronized H2 and external structural computations.
The earlier matched-baseline performance bottleneck is no longer an open
execution gate in this build. Thread settings and dependencies are documented.
The quick release verifier is distinct from this complete scientific replay.

HISTORICAL D APPROXIMATELY 3
The recovered historical coding assigns Cys second-count 1 and effective
Arg charge 0.5. Its template distance is 0.017693128924; the supplied
five-million-draw replay gives zero hits and plus-one p=1/5000001.
The descriptor/constant design was response-exposed. This is conditional
fixed-basis evidence, not a selection-adjusted dimension test. Correcting
only Cys to N/O count 0 changes distance to 0.0774229837224 while improving
unconstrained fitted R2. The historical p-value is not transferred to the
corrected design. Detailed treatment is confined to SI S12/S22.

TEXT AND LAYOUT
The paper centers on distinct preservation targets, with direct q2 and
singleton-Arg failure certificates, fiber losses, regularization scope,
shared-profile testing, external rank demonstration and applicability limits.
Speculative mechanisms and the full claim ledger remain principally in SI.
The bibliography has 36 entries, with the targeted 2021--2026 literature
comparison in S24. No exhaustive systematic priority claim is made.

All 19 manuscript PDF pages and 25 SI PDF pages were visually inspected.
Editable Word renderings were separately inspected: 19 and 24 pages.
The cover letter is one page. Pagination is deliberately format-specific.
There are six main sections, S1--S26, C01--C49, equations (1)--(16) and
(S1)--(S16), three main tables and nineteen SI tables. Four main figures,
eight diagnostic figures and one new SI figure have PDF/SVG/PNG masters.
All eight regenerated diagnostic rasters match the previously verified
release pixel-for-pixel. A truncated diagnostic PNG was regenerated before
packaging; all PNG streams now fully decode. The five publication-facing
figure masters were inspected in the documents.
Machine layout checks: 31/31 PASS. Fonts are embedded; no missing-glyph,
overfull-box or oversized-float warnings. Equations, tables and figures
have separate consistent numbering. No comments or tracked changes exist.
Empty Word comments containers carry no review content. PDF and DOCX are
built from the same Markdown text, with included editable sources/build code.

INTEGRITY AND LIMITS
Both SHA-256 manifests are rebuilt after all edits. The root manifest covers
the nested predictive manifest as well as scientific and publication files.
Final verifier logs and external fresh-extraction evidence accompany release.
The scientific ZIP embedded in the submission package is byte-identical to
the separately supplied scientific ZIP. SHA-256 sidecars identify both.

No new wet-lab or prospective response data were created. Noncanonical
physical transfer and crossed Born/support experiments remain OPEN, as do
Hessa/Moon uncertainty provenance and panel-specific power calculations.
The original prospective decision rule is preserved; strengthened planning
requirements do not silently replace it. Chemical extrapolation failure,
Moon conditional status, limited A20 support and model-identity instability
remain visible. No journal submission or public deposition was performed.
