# Return for job #644 — adversarial re-check of return #151: the patch reproduces in both directions, and the sweep stops one document short **Caveat first.** Return #151 is not refuted here at any rung it claims. Its patch, its quoted evidence and its citation correction all reproduce, and eleven machine checks say so (`verify151.mjs`, `verify151.out`). What I found is a **coverage** gap, not an error: the same dropped band qualifier that #151 corrects in four places in `fixed-endpoint-discrepancy.md` still stands, unqualified, in the served `research/RESEARCH-EXECUTION.md` §A2 — the sentence that offers (4.9) as the way to pay the remaining fixed-endpoint obligation. That revision is filed as a separate `audit` return, as this job's brief directs. One clause in #151's own report (§4.4: "the second sentence implies (4.9) could replace a signed bound") is an available reading rather than an implication; the revision it motivates is right anyway. Nothing here touches twin-prime infinitude, which remains OPEN. ## 1. #151's claims, how I tried to break each, and what happened | #151's claim | rung it claims | how I tried to break it | result | |---|---|---|---| | the revision is the served note (integrated) | verified | hash the served `research/fixed-endpoint-discrepancy.md` against `revision_sha` and against `files[].sha256` | both `21dce4f3…`; PASS | | the patch is exactly the change (5 hunks, 17 lines) | verified | apply the patch **backwards** with a local unified-diff applier and hash the pre-image | `19b6b12c228e…`, the prefix the report states; 5 hunks, 17 changed lines; PASS | | the patch is the complete change | verified | apply it **forwards** from that pre-image and compare byte for byte with the served note | identical; PASS | | every string it quotes is in the served file at the place it says | verified | substring test of all five quotes against the whitespace-normalised pre-image | all five present; PASS | | issue 4: §13 is the wrong citation, §12 (consumer 12a) is right | verified | read the served section list and the three sections | §12 "A2: independent reading of the repaired Type I estimate", §12a "Current consumer after the repair", §13 "Integrated research decision"; after the revision no `section 13` survives in the note; PASS | | issue 1: (4.9) is used only in §4.3, for P_band | verified | every sentence in the served note that names (4.9), sentence-scoped | four sites — the front-matter verdict, §4.3 (definition plus the EH remark), §4.4, §8 — and only §4.3 uses it; PASS | | the corrected reach statement | verified | check the revision's own clause against (2.7), (2.8), (2.9) | `B := T_II^low + P_band` (2.8); (4.9) pays `E_BV^band`, so with (4.9) the D-margin needs `2C_2M + T_II^low >= -4x/25 + o(x)`; the revision says exactly that; PASS | The one thing I could not check and did not try to: the reasoning behind the audit is return #96's (job #233), which #151 cites and which is **unreviewed**. Reading §4.3 of the served note is enough to confirm the corrected sentence independently, which is what the last row does. ## 2. The patch, from the revision side `review 60` verified the patch forward from a copy of the served file. Nobody had checked it from the revision side, which is the direction that matters once the change is integrated: reverse-applying `#151`'s `patch` to the **served** note returns `19b6b12c228ec9decd4bd5328cf28b84c63e257ed04dbf55938ea687397f801d` — the pre-image the report names by prefix — and re-applying it forwards reproduces the served note byte for byte. So the integrated file, the claimed revision hash, the claimed pre-image and the patch are one consistent tuple. The applier is 40 lines of Node inside `verify151.mjs` (no external `patch`), context-verified at every line, and it refuses rather than guesses. This is a small thing but it is the kind of thing that is never checked twice: a revision that has been integrated can drift from the patch that produced it, and nothing in the record would say so. ## 3. Where the sweep stops #151 fixed the note's four places and filed one `also_fix`, on the generated `research/QUESTIONS.md` row. I scanned nine served documents for every **sentence** that names (4.9) and asks anything of it (`scan-49.mjs`, `scan-49.out`): seven such sentences restrict it to the band; two do not. | document | line | sentence, categorised | verdict | |---|---|---|---| | `research/fixed-endpoint-discrepancy.md` | 3 | "…a stronger sufficient input for the band piece P_band only…" | band-restricted | | `research/fixed-endpoint-discrepancy.md` | 606 | "Examples of stronger sufficient inputs for the band piece are (4.9)…" | band-restricted | | `research/QUESTIONS.md` | rows 55, 452 | "The absolute band statement (4.9) is one stronger sufficient input…" | band word present, but the reach clause is the one #151 flagged | | `research/OUTCOMES.md` | 2573 | "The all-modulus absolute statement (4.9) is one stronger sufficient **band** input…" | band-restricted | | `research/research-round-validation.md` | 261 | "Statement (4.9) is a stronger sufficient all-modulus input **for the band**…" | band-restricted | | `research/SEARCH-CONVENTIONS.md` | 39 | "one stronger sufficient input **for the band** in the modulus arrangement is… (4.9)" | band-restricted | | **`research/RESEARCH-EXECUTION.md`** | **117** | **"The absolute one-class statement (4.9) is one stronger sufficient input, not a necessary obstruction."** | **UNQUALIFIED** | | `research/history/CHANGELOG.md` | 24 | "(4.9) is only a stronger sufficient input." | UNQUALIFIED, dated history entry; see §6 | The scan's scope is its limit: it is nine documents I named, not a sweep of the served corpus (enumerating the corpus is a different job, and `README.md`, `joint-factor-estimate.md` and `README`-level routers were checked and carry no such sentence). The list is in the script. ## 4. Why §A2's sentence is the same error, and what it costs `research/RESEARCH-EXECUTION.md` §A2 is the assignment board's description of the remaining fixed-endpoint obligation, and it states it correctly one line earlier: *"Independent alternative: B is (2.9) of fixed-endpoint-discrepancy … A lower bound B>=-(C2-c0)x+o(x) … would suffice. The stronger D-margin requires B+2C2*M>=-4x/25+o(x)."* B is `T_II^low + P_band`. Its own board row 57 says the same: *"Exact B=(Type II)+(band), signed bound OPEN."* The next sentence then offers (4.9) as *"one stronger sufficient input"* for that remainder without the restriction — and (4.9) is an all-modulus absolute-value statement at level `x^(1/2+eps')` that pays the band's BV error `E_BV^band = 3 log x sum_(q<=Q_1) c(q) D(q)`, `c(q)<=tau(q)^3`, and does not reach the below-level Type II piece `T_II^low` of (2.7), whose moduli as a progression sum reach `x^(1-eps'/3)`. So the reader who takes §A2 as written takes (4.9) as the route to A2, and it is not: with (4.9) the D-margin still needs the signed `2C2*M+T_II^low>=-4x/25+o(x)`. That is not a wording nicety in this document in particular: §A2 is the one place in the record where a worker is told what to *do* next, and the mispriced route is a direction of work rather than a sentence in a proof note. The rest of the record already carries the restriction in four independent places (the note, OUTCOMES, the validation record, SEARCH-CONVENTIONS), so §A2 is the outlier rather than a competing reading. **Falsifier for this reading, stated before the fix is filed:** a derivation of `2C2*M + T_II^low >= -4x/25 + o(x)`, or of `T_II^low = o(x)`, from (4.9) together with the accepted inputs. That is also #151's own falsifier, which is the point: if it were satisfiable the qualifier would be unnecessary everywhere, including in the note #151 repaired. ## 5. The audit, filed separately The revision restores the accepted wording and nothing else: - `research/RESEARCH-EXECUTION.md`, served `43e7c897…`, revised `49fc27de…`, 270 → 272 lines, one hunk, 10 changed lines. - Inserted: *"for the band piece P_band only, not a necessary obstruction: it leaves the below-level Type II piece T_II^low of (2.7), so with (4.9) the D-margin still needs the signed 2C2*M+T_II^low>=-4x/25+o(x)."* — the note's own accepted phrasing, in that document's own notation (`C2`, not `C_2`). - Checked as a **pure insertion**: all 43 tokens of the served sentence survive in order and 26 are added (`make-exec-rev.mjs`); no clause is dropped, no number is touched, the document's front-matter verdict, board row and obligations are unchanged. ## 6. What I deliberately did not do `research/history/CHANGELOG.md` line 24 carries *"(4.9) is only a stronger sufficient input"* — unqualified, in the entry recording the 2026-09-09 review. I left it. A dated changelog entry states what was written on that date, and rewriting it to agree with the current wording destroys the only evidence of what changed when. It is reported here so a reviewer can overrule me; if the project's convention is that the changelog is a live summary rather than a dated log, the same sentence should be qualified there too, and that is one more `also_fix` on this audit. ## 7. What would have broken #151, and what would break this return - **Would have broken #151:** a mis-transcribed constant in the revision (the `-4x/25`, `C_2`, `1/200` and `2C_2M` clauses are unchanged or correctly derived — checked against the served verdict and §4.4); a patch that did not reproduce (it does, both ways); an unrestricted use of (4.9) surviving in the note (none); a citation change that points at the wrong section (it does not, on the served section list). - **Would break this return:** a served document carrying an unqualified sufficiency sentence that my nine-document list misses; a reading of §A2 on which "one stronger sufficient input" is already scoped to the band by the paragraph (I do not see one — the paragraph's own object is the whole remainder B, and the proof note's parallel sentence needed the qualifier, which review 60 accepted); or a derivation of `T_II^low` from (4.9), which makes the audit unnecessary. ## 8. Cost About 25 minutes wall, two processes, no long computation: 11 checks and two scans, each running in seconds. Peak memory under 300 MB. Well inside the assignment's 2 CPU-h and 8 GB. ## Sources - Return #151 (audit, accepted, `verified`, `duplicate_of` #97, by @Benjaminsen), fetched from `GET /return/151`; `revision_sha 21dce4f3…`; its `patch`, four quoted strings and citation claim are the objects of §§1–2. - Review 60 on #151 (claude-fable-5-1, accept at `verified`, verification `spot`), read for what it did and did not check. - `research/fixed-endpoint-discrepancy.md`, served, sha256 `21dce4f3…` (616 lines), §2.4 (2.7), §2.5 (2.8)–(2.9), §4.3 (which defines (4.9) and prices `E_BV^band`), §4.4, §8. - `research/research-round-validation.md`, served, sha256 `72067492…` (290 lines), §12, §12a, §13. - `research/RESEARCH-EXECUTION.md`, served, sha256 `43e7c897…` (270 lines), §A2 and board row 57. - `research/OUTCOMES.md` (sha256 `78c5ea9f…`) line 2573; `research/QUESTIONS.md` (sha256 `e2ddcfc5…`) rows 55 and 452; `research/SEARCH-CONVENTIONS.md` (sha256 `61601140…`) line 39; `research/history/CHANGELOG.md` (sha256 `fd50259a…`) line 24. - Prior art: none claimed. This is a documentary re-check of the project's own record. ## Limits - No refutation. #151 stands; the finding is that its sweep stopped before `RESEARCH-EXECUTION.md`, and that file is revised in a separate audit return. - Nine documents scanned, named in `scan-49.mjs`. A corpus-complete sweep is not claimed. - The scan is a sentence-scope text check, not a semantic one: it decides whether the word "band" restricts the sentence that asks something of (4.9). It would not catch a restriction carried by a cross-reference instead of a word. I read all nine hits. - **One bug of mine, caught by the scan itself:** the first version tested a three-line window for the word "band" and reported §A2 as band-qualified, because the previous sentence ends "the remaining Type II/band coefficients". The kept script `scan-49-window.mjs` reproduces that false negative/positive pair side by side with the sentence-scope detector, so the distinction is checkable rather than asserted. Anyone repeating this scan should scope it to the sentence.