FINAL-FILE REVIEW SUPPLEMENT Finite Noisy Audits with Learned Baselines 9 November 2026 (Asia/Bangkok) Two separate internal AI review passes found no supported fatal, major, or required minor mathematical correction in the exact final source. Source: https://www.b-theory.com/manuscripts-html/finite-noisy-audits/manuscript.txt SHA-256: 139fd3e6b149b5338e380a2c9479beae0a00b655bd5b9985a641df26f183ecd7 Complete source coverage: 167 lines; 20485 bytes. Reviewed release commit: cf93a3a59db0b587dd45f524f7f56da08c0ec190 METHOD AND SCOPE One proof review and one adversarial review, each based on an actual complete-file read and independently computed hash. Independent here describes separate review runs, not external institutions or human peer review. No mathematical source revision was required. Original source, PDF, public archives and historical review records remain unchanged. This dated supplement adds final-file review evidence; it does not rewrite the earlier review history. PROOF REVIEW No required mathematical revisions. Independently reconstructed rowspace identifiability, separate constant-risk Bayes/minimax formulas, uniqueness-based all-controller memory lower bounds and packed attainment, simulation/encoding arithmetic, and randomized minimax risk for the specified unseen defect. Exact rational arithmetic confirms reported risks and costs. Every majority time-cut residual width was additionally enumerated for odd lengths 1 through 31; this finite check corroborates the general analytic proof. Prior requests: All seven documented required concerns are resolved: fixed allocation versus global optimization; calibration versus family validation; explicit interfaces and all-controller lower bound; enlarged-family/fresh-coin contract; simulation versus architectural superiority; representable header limit; and conservative benchmark contribution. The earlier optional estimator-naming suggestion is not required for validity. Limits: No independent execution or byte-equality certification of the supplied ROMs. No PDF or remote-deployment verification. Results remain conditional on the supplied structural equalities, independent noise, fixed blocked clocked schedule and stated losses; they do not certify unknown defects or minimum physical hardware. ADVERSARIAL REVIEW No supported required mathematical change 1. Checked identifiability quantifiers: failure of rowspace containment produces indistinguishable mechanisms with a differing target; repeated observed basis rows give uniformly reliable recovery only asymptotically. Finite noisy sample zero-error identification is not asserted. 2. Rederived the independent (theta,alpha) coordinate prior through the bijection, unique odd-block posterior majorities, XOR posterior-bias sign, independent diagnosis groups and vector/full-mechanism MAP rules. Constant error probabilities for each fixed mechanism match the uniform-prior Bayes lower bound separately for each declared loss. 3. Independently recomputed exact rational binomial tails and all five risks at (n,r)=(3,3),(3,13),(5,7). Recovered H=.104, V=.280676864, p=.186368, G=.338002968576, M=.584602309832559 at (3,3), and every rounded Table 1 risk and ROM byte count (117805 and 73645). The base ROM formula yields 22689 padded bytes. 4. Attacked the memory lower bound against alternative controllers: every transcript/query has positive probability and unique Bayes outputs, so matching the optimum cannot trade errors among transcripts. Across possible late queries, output functions recover all required majority decisions. Common future words distinguish each prior decision vector and residual majority function, even when a changed theta1 output is exposed only through gamma. 5. Independently enumerated exact-majority residual threshold functions at every cut for n=1,3,5,7,13,19, including the n=1 edge case; peak width is (n+3)/2. Multiplication by retained prior decisions gives the two stated maxima and terminal capacities. Time-indexed reuse of state labels makes the maximum, not sum of widths, sufficient. The external clock and readout/log restrictions are essential and explicit. 6. Checked compiled simulation is induction on full persistent configurations, not independent architecture discovery. The dense bit formula gives 181505 logical bits at (3,3), and is explicitly bounded by 16-bit-header representability. 7. Rederived the enlarged-family bound by counting twelve weighted response positions: eight baseline, three audit, one conditionally fair hidden target. Exact rational recomputation gives 137/1000. Fresh independent random output yields constant risk for both delta values; nominal adverse risk is .17. The deterministic ROM is not claimed to implement the coin and no arbitrary-defect guarantee follows. Prior requests: After independent evaluation, read adjacent responses-to-reviewers, ai-review-summary and source-proof-ledger, and the concluding original proof/hostile reports. All required scope/precision requests are present: audit is calibration rather than structural validation (9,32); allocation is fixed and not globally optimal (96,106); all-controller memory uniqueness bridge (77); separate interfaces and public clock (34-36,71-83); 16-bit-header restriction (94); enlarged-family fresh randomness (110-121); matched simulation and calculated versus generated alternative ROMs (87-106). No prior required mathematical request remains open. Limits: This pass did not inspect the ROM bytes, execute the delivered large transition enumeration, inspect the PDF, certify the archive/manifest, or re-retrieve cited literature. Arithmetic and majority-function checks were independently implemented, but finite checks do not replace the analytic all-n,r arguments. No human peer review, physical validation, general optimal-allocation result, or historical-priority claim is made. OVERALL LIMITS Internal AI review is not a guarantee of correctness, human specialist peer review, proof-assistant certification, historical-priority clearance, or physical validation. These source-review passes did not repeat PDF layout or full software-package certification. Earlier production checks and current publication verification are separate. UNCHANGED PDF https://www.b-theory.com/manuscripts/finite-noisy-audits-v1.0.pdf SHA-256: 9439ad244b8b71360d5abf4a5124ac4f2d393b0bd3116c515e74c754575bca00 A SHA-256 match binds review to file bytes; it does not itself prove mathematical correctness. Original review and publication-package records retain their historical scope.