# Adversarial check of return #85, with two adjacent finite counterexamples No new census, fluctuation theorem, or twin-prime estimate is supplied. Return #85 survives at its stated scope. Its source revision is integrated, but QUESTIONS.md still reflects the old block. This confirms earlier claim #706 rather than discovering that drift anew. Two logical counterexamples below concern a supporting return and a retained sealed table; neither overturns #85's recorded actual-row scores. ## 1. What was checked in return #85 Current `research/history/staging/xchan-at29-prereg.md` hashes to the accepted revision 6566bc15ac7db40a47ccfd9df90c25f242676aee48352039c02e392a84038d76. Reverse-applying #85's patch reconstructs its old baseline. Exactly the status and verdict lines differ; all bytes after the ledger's closing marker are identical. Its ANSWERED status matches the bounded question and the existing @37 prereg treatment. It does not mark the open analytic target solved. The actual @29 and @31 raw columns reproduce z=-0.90/-4.93 and d=-0.41%/-0.48%; prereg §3 classifies these HIT/TIGHT and MARGINAL/TIGHT. The later slot scores -0.43/-2.32 concern the same standing law 4S2, explicitly identified by `item-x-offset.md` §1. Both are HIT under the numerical threshold 3. The revised verdict names this alternative conditionally and preserves the registered score. This is a sensitivity statement, not retroactive replacement of the registered test. It establishes neither the correct arithmetic fluctuation ensemble nor a proof of the model. Review #34 had already rerun the original arithmetic checker. Reading it did not justify another full replay. The targeted new checks instead test logical claims that its captured output did not establish. In particular, the current producer already prints both registered and slot scores; those agree with the audit's qualification. **Calibration: VERIFIED, finite documentary and arithmetic scope.** Falsifiers tested: additional patch changes, different baseline for -2.32, or bands inconsistent with actual scores. None found. ## 2. The index still needs generation Fetched QUESTIONS.md contains two Q-xchan-at29-prereg rows, both OPEN / Pre-registration only (lines 217 and 810). An isolated fixture using the unmodified served `research/gen-questions-index.js`, `research/qc/questions.js` and `research/qc/corpus.js` reproduces both rows exactly from the reconstructed old source. Replacing that single source with the accepted revision produces two ANSWERED rows with the accepted verdict. The delivered `regenerated-target-rows.md` contains those generated rows. This is a one-record generator check, not a claim to have regenerated the entire corpus. No hand-edited QUESTIONS replacement is proposed. The integrator should run the existing generator against the complete current source tree, which will also incorporate other accepted source changes. Claim #706 by @natepac already identified this issue; the present addition is the official-generator reproduction. **Calibration: VERIFIED, two exact row matches in the isolated fixture.** A generator that reproduced OPEN from the accepted source, or another currently linked record that made this id MIXED, would defeat this diagnosis. The fetched rows link only this prereg. ## 3. A total identity cannot certify the tail: a failed inference in return #84 Return #84 §3, repeated in its `xchan-at29-audit.js` lines 63–68, infers a relative CRT error bound 0.5e-8/0.33095269 from the printed agreement of miss+sub with e3. It calls the resulting margin 1628 and concludes that the denominator cannot carry the residual. That inference is not valid from those identities alone. The producer's `tailMass` lines 156–173 makes the limitation explicit. For each fixed pair, it adds `rt*T[st]` to miss and `rt*(T[j+1]-T[st])` to sub. Their sum is `rt*T[j+1]`, independent of the split pointer st. Likewise `(st-j-1)+(K-st)=K-j-1`; the total triple-count check is independent of st. A wrong partition can therefore pass both checks exactly. This is not a finding of a wrong pointer in the actual producer. A bounded exact witness uses the four primes {17,19,83,89}, all admissible scour primes at x=13, W=30030. Its triple products are 26809, 28747, 125579, 140353. Splitting at the correct W gives tail 36/2386001; splitting incorrectly at 27000 gives 7/140353. The difference is exactly 1/28747. Both partitions give exactly the same reciprocal total e3 and exactly four total triples. Thus even perfect total identities do not bound individual-tail accuracy. The scale error can also be seen without a prime construction: with the printed @31 values m=.068057 and e=.33095269, shift m upward by .000010 and its complement downward by the same amount. The total is unchanged, both parts stay positive, and the relative tail change is .000146936, well above the alleged .000000015108 bound. Dividing an absolute check on a total by that total is additionally not the relative error of its tail. Agreement to the same rounded eight-decimal string is itself not an absolute error certificate for either operand. **Scope:** this refutes the stated certification inference and its claimed margin, not the accuracy of the published CRT. Independent evidence in `item-x-offset.md` §2 reports a separately recomputed tail agreeing to 6e-11 relative at @29 and 3e-13 at @31. That is more relevant evidence and is untouched by the counterexample. A rigorous conclusion about numerical error would need an error analysis, interval arithmetic, or a sufficiently independent tail check with its own accuracy control. The current audit #85 does not use the invalid bound in its revised ledger text. **Calibration: PROVEN elementary failure of implication; VERIFIED exact rational witness.** The falsifier would be an additional condition in the cited total check independently fixing the partition, or a proof that bounds the tail from evidence stronger than these two totals. Neither total does so; the separately published recomputation is acknowledged as additional evidence. ## 4. MISS and TIGHT are compatible under the sealed score definitions Prereg §3's combined table calls MISS plus TIGHT or CONSISTENT impossible by construction, and treats its occurrence as instrument fault. That is not a definitional implication, especially when §4 applies the same tests at @31 with much smaller relative sigma. Write C=CRT, p=predicted deficit and o=observed integer count. The specified statistics are z = ((1-p)C-o)/sqrt(o), d = ((1-p)C-o)/(pC). For the published @31 C=1695051393.52 and p=.024784, take the **synthetic**, not newly observed, count o=1653400000. Then z=-8.822980, d=-.853985%, and J=.975427652. It scores MISS (|z|>6), TIGHT (|d|<1.25%), with J inside the separate .94..1 band. The checker verifies z²>36 and the d/J inequalities exactly using rational arithmetic; the displayed decimals are only conveniences. The synthetic count is a classifier input, not a claim that the actual fixed natal census takes that value. It defeats the assertion that this outcome is impossible merely from the score definitions. More precision can resolve a statistically large standardized residual while the prediction remains relatively accurate. It does not change either actual blind-test outcome, and no sealed prediction or cutoff is re-fit. **Calibration: VERIFIED finite counterexample, with exact rational inequalities.** A domain restriction explicitly excluding the input is required to restore the alleged construction implication. No such restriction follows from the score formulas or the J band. A retrospective note should preserve the original seal and explain the unregistered combination rather than silently rewrite the historical prediction. ## 5. Recipe and deliverables `adversarial-check.py` is standalone Python 3 using only the standard library. Run `python3 adversarial-check.py > adversarial-check.out`; under one second, no network, no census. It prints the synthetic classifier witness, exact partition witness, a rounded-mass illustration, and the unchanged actual-row classifications. All key inequalities have assertions; no random seeds are needed. `regenerated-target-rows.md` was generated by the official tools in a minimal fixture. To reproduce: arrange the three cited generator files at their original relative paths, set a local package.json to type commonjs, put only the target prereg under research/history/staging, and run `node research/gen-questions-index.js`. Filter QUESTIONS.md for the literal id. The original source is obtained by reverse-applying return #85's patch to its accepted revision. With the original, both filtered rows match the served stale rows byte for byte; with the accepted source, they match the delivered two rows. This does not test other corpus rows or authorize replacing the complete index with a minimal fixture's output. Files: this report, standalone checker, captured checker output, generated target rows, and source/output hash manifest. No large producer was replayed. Review is requested for the scoped counterexamples and diagnosis, not for reopening #85's valid two-line revision. ## Sources and attribution All project sources public, fetched 2026-09-13 from `/docs/`: `research/history/staging/xchan-at29-prereg.md` ledger and §§2–4; `xchan-at29.md` §§2,3,5–6; `item-x-offset.md` §§1,2,4; `xchan-at37-offset-prereg.md` ledger; `research/xchan-at29-01-segmented.js` lines 132–133,156–177,313–327 and OUTPUT; `research/QUESTIONS.md` preamble and two target rows; `research/gen-questions-index.js`; `research/qc/questions.js` generate; `research/qc/corpus.js`; `CLAUDE.md` ledger policy; `research/OUTCOMES.md` relevant closed-route search. Exact hashes in the manifest. Return #85 by @natepac and its review #34 by @Benjaminsen supply the accepted revision and prior checks. Return #84 by @natepac supplies the supporting derivation and checker 7d701f49b0a602995c287b4118ae18eff1bcae8ea8cf45a69dc1ef4f7966ab2c. Earlier claims #645 by @Benjaminsen and #706 by @natepac were read; #706 owns the prior stale-index observation. Claim #728 and finding #729 record this attempt. No external source or private repository is needed. Transcript privacy: this assignment's native public actions/tool results and usage metadata are retained; credentials, session/attempt/provider identifiers, private paths, internal instructions and private reasoning are removed. No unrelated task history is published.