FINAL AUDIT REPORT
Protein interaction-response release v1.0.3
Audit date: 15 September 2026
Manuscript: From Microscopic Ensembles to Validation-Safe Transfer-Energy Compression

OUTCOME
The scientific replay completed in a separate clean working copy using only
included numerical inputs. Its terminal result was COMPLETE SCIENTIFIC REPLAY:
PASS. The exact mathematics, finite certificates, empirical predictions and
remaining inference boundaries are distinguished throughout the manuscript.
The release is self-contained; no unpublished companion work is required.

FRESH COMPUTATIONAL CHECKS
- Interaction geometry: 36/36 checks PASS, including the 46-entry claim registry.
- Structural enumeration and matched Ridge/PCR/PLS: 24/24 PASS.
- Core H1/H2a and chemotype replay: 12/12 PASS.
- Charge-magnitude comparator sensitivity: 5/5 PASS.
- Claim-contract mutation controls: all 13 invalid cases REJECTED.
- Exact representation certificates: 3/3 PASS. All 1,023 D10 subsets enumerated;
  minimum A20-injective encoding has five coordinates, with 16 such subsets.
- H1 adaptive null: all four full 50,000-permutation campaigns matched their
  reference counts, quantiles and observed selector identities.
- Fixed-class stress null: all four full 50,000-permutation campaigns replayed.
- WW bounded-error sensitivity: two full 10,000-perturbation campaigns matched.
- H2 alignment: two full 10,000-draw campaigns replayed against the fully
  specified current reference, including both zero counts and all quantiles.
- H2 local sensitivity: 9,000 simulations, using the complete S10 protocol.
- H2 vectorized operators: 2/2 independent explicit-refit controls PASS.
- Quantitative fiber and identifiability certificates: 21/21 PASS.
- Historical Cys coding contrast: 2/2 PASS.
- Historical fixed-basis null: all five million permutations recomputed by an
  independent rational-OLS implementation; 4/4 comparison groups PASS (six
  event counts, five observed statistics, three minima, sixteen quantiles).
These are separate verification suites, not independent discoveries or new
experimental replications. The full execution transcript is supplied in
verification/COMPLETE_SCIENTIFIC_REPLAY.log.

REPRODUCIBILITY AND SCIENTIFIC INTERPRETATION OF THE HISTORICAL RESULT
The fixed historical design reproduces coefficient distance 0.017693128924
from the selected (2/3, -1/3, 3) template. There are zero hits in five million
response-label permutations, giving plus-one p = 1/5,000,001. The executable
protocol, input CSV, random seed, draw order and reference outputs are included.
This resolves computational reproducibility on THAT design. It does not repeat
response-exposed descriptor, effective-Arg or constant selection, and is not
selection-adjusted evidence of intrinsic physical dimension.
Historical Cys hydrogen-bond coding is 1; chemically exact N/O count is 0.
Changing only this entry raises the template distance to 0.0774229837224 while
fitted R-squared rises from 0.96206031 to 0.96578749. The historical p-value is
not reused for that changed design or for the physical q2 representation.
The plus-one p is not a confidence bound. The separate one-sided 95% zero-hit
binomial upper bound is approximately 5.99146e-7 under the fixed conditional null.

H2 SIMULATION PROVENANCE
The observed aggregate Q2 values remain 0.5490920693730571 and
0.5756109186180024. A complete alignment-null draw order is now specified and
executed. Both current campaigns have 0/10,000 exceedances. Current median/
95th/99th-percentile values are, respectively:
  four endpoints: -0.271259357911, -0.0658844667974, 0.0108555046645;
  three endpoints: -0.284347788611, -0.0584938534760, 0.0297975331223.
The earlier summary records did not expose sufficient producer detail to
recover their exact simulation stream. Their quantiles are superseded by this
explicit recomputation; agreement of zero counts is not claimed as byte-for-
byte recovery of the earlier simulation. A second replay verifies the current
fully specified stream.
The local sensitivity experiment is a newly explicit, response-conditioned
illustration, not recovery of the earlier calibration. At signal R2 of
0.3/0.5/0.7, near-shared exceedance counts are 42/54/49 and separate-signal
counts are 112/111/129, each out of 1,000. Shared calibration gives 50/1,000 by
construction. Low simulated separation does not prove physical nonidentifiability.
Diagnostic Figures 5 and 6 were regenerated from these current ledgers.

PUBLICATION AND LITERATURE
The manuscript contains 33 references, with a targeted 2021--2026 comparison
in SI S24 and an explicit separation of overlap from the integrated contribution.
Coverage includes amino-acid scale organization, contextual transfer energetics,
membrane insertion, molecular/protein representations, and similarity-aware
validation. This is not an exhaustive systematic review of all protein research,
and it establishes no unmatched benchmark superiority or blanket priority claim.
Thermodynamic-cycle algebra, projection identities and cospark are acknowledged
as established foundations. The contribution is their integration with observer
contracts, quantitative refinement and validation-safe model spaces.
References, author identity, corresponding email and cover letter are aligned.
The main manuscript uses double body spacing and continuous line/page numbering,
consistent with the journal's Methodology submission guidance checked on the
audit date. The Scientific Contribution abstract subsection has three sentences.

LAYOUT AND DELIVERY CHECKS
All 22 manuscript pages, 24 SI pages and the one-page cover letter were rendered
and visually inspected. All four publication figures and eight diagnostic
figures were inspected. Diagnostic Figures 3--8 were regenerated from current
machine-readable outputs; the input hashes are recorded. Native editable Word
equations are retained; comments and tracked changes are absent. Two SI citation
numbers and a malformed certificate path were corrected after visual review.
The publication PDFs correspond to the final DOCX renderings.
Release integrity, predictive contracts and independent representation checks
are run by verify_release.py. Both manifests and external ZIP SHA-256 sidecars
are regenerated after final synchronization. Archive checks include fresh
extraction, nested release identity and byte agreement of duplicated documents.
The extraction verifier's execution record accompanies the delivered audit.

REMAINING SCIENTIFIC BOUNDARIES
- Moon-Fleming remains conditional, with fixed-class stress p = 0.1127177456.
- Hessa and Moon row-level uncertainty gaps remain open.
- WW-octanol unseen-chemotype stress gives Q2 = -1.9766598014.
- Shared prediction does not identify a unique common physical generator.
- Finite injectivity and structural rank do not identify intrinsic response rank.
- Physical-carrier tests and new blind prospective outcomes remain open.
No new experimental endpoints or prospective outcomes were invented. No journal
submission, external public deposition or repository DOI assignment was performed.

LICENSES
Original code: MIT. Original documentation, figures and curated data: CC BY 4.0;
underlying published observations retain their cited provenance.
