Independent strict review of M15-v1, M24-v1 and M05-v1
22 October 2026; literature cutoff 22 October 2026

All three frozen candidates are supported internally, cover their exact originals, and have a precise contribution beyond the primary sources checked. No decisive proof gap, counterexample or known full prior coverage was found. This is a fallible independent agent review, not human peer review, external acceptance or a worldwide priority certificate.

Separate judgments:
M15-v1 — validity supported internally; complete original; difference supported; external acceptance not established.
M24-v1 — validity supported internally using the named imported affine theorem; complete original; difference supported; external acceptance not established.
M05-v1 — validity supported internally; complete original in its specified input model; difference supported; external acceptance not established.

What is actually new:
M15: global L1 error regrouping and a positive, polynomially normalized moment envelope give a bounded polynomial sampler. Point-probability approximation alone is inherited and insufficient.
M24: the generic 1:m multiplier/saturation argument supplies a rank-two EPR witness on one extractable Bob mode. The external two-mode affine inequality and Gaussian covariance roof remain inherited.
M05: the fully charged recursive implementation, correction oracle, simulator inputs and classical assembly close the cost analysis of an already proposed approach. The architecture itself should not be presented as new.

Qualifications that must survive drafting:
M15 fixes depth d and peak exponent k, permits inverse-polynomial error, and has very large polynomial exponents. The heavy output is not supplied. Its proof can be instantiated at epsilon=n^(-k), so an unnecessarily long rational encoding of a larger requested epsilon need not be read for this fixed-promise guarantee.
M24 treats finite one-versus-many systems, including arbitrary q/p correlations, physical boundaries and unrestricted decomposition measures. It relies on the checked external affine entropy theorem; the review did not reconstruct that source's entire proof.
M05 displays exp(C x^3) N (1+b)^c Q^(C x+c), including input-bit work and an O(x) logarithmic exponent. It proves state preparation, not whole-global-semigroup simulation.

Evidence:
The original and entire frozen proof were read for each candidate, together with relevant imported interfaces. Prior internal verdicts were not used as evidence. Fresh targeted controls probe the most vulnerable steps: actual support truncation and rejection for M15; strict-physical KKT/rank/witness algebra plus a boundary negative control for M24; and a full matrix amplification recurrence for M05.
There were 18 clean supplied-checker runs plus two clean replays of the fresh controls. Every run exited successfully; corresponding stdout was byte-identical. Exact rational controls cover M15/M24 finite algebra. M05's fresh control is floating-point diagnostic and is identified as such. None of these finite calculations replaces the analytic proofs.
All 267 frozen input hashes still match. Only this independent review directory was written.

Literature gate:
source-comparisons.txt and source-comparisons.json identify 12 primary-source comparisons, with exact theorem/section/equation locators and URLs. They distinguish imported/covered components from partial overlap and the actual additional claim. Current source versions were checked where material. There is no specific unresolved prior-coverage lead or dependency mismatch remaining in this review. Ordinary non-exhaustiveness is the remaining search limitation, not a concrete reason to withhold these three.

Artifacts:
M15-review.txt / M15-review.json
M24-review.txt / M24-review.json
M05-review.txt / M05-review.json
source-comparisons.txt / source-comparisons.json
frozen-inputs.json
integrity-and-reproduction.json / replay-summary.json
targeted-controls.py / targeted-controls.json
replay-1/ and replay-2/

Next justified action:
Retain these unchanged candidates for the parent's combined release assessment, with the qualifications and attribution above. This review makes no judgment on the other seven candidates and does not satisfy the user's all-ten condition by itself. It performs no submission, publication, catalog update, repair or additional research round.

