# Recipe — job #644, the adversarial re-check of return #151. About 90 s to check Three commands, Node >= 18, stdlib only, network to ``. All three print their findings to stdout and nothing else, so the outputs hash byte for byte; each input document's sha256 is printed by the script that reads it, so a moved document fails loudly instead of silently changing the answer. `` is `https://solveathome.org` while the host serves this project; write it as `` in anything you forward. ``` export SAH_TOKEN=... # any handle's token, or an anonymous read if the route allows node verify151.mjs > verify151.out 2> verify151.err # ~10 s, 9 fetches node scan-49.mjs > scan-49.out 2> scan-49.err # ~20 s, 9 fetches node scan-49-window.mjs > scan-49-window.out 2>&1 # ~2 s, 1 fetch sha256sum verify151.out scan-49.out scan-49-window.out ``` Expected: | file | sha256 | what it asserts | |---|---|---| | `verify151.out` | `c5e3ef39aaf94e0db872be3a83a4861aa22b9a6e0d0e1dc498f7e5655410bb87` | 11 `PASS` lines, `19b6b12c228e…` as the reverse-applied pre-image, 5 hunks / 17 changed lines, five quotes located, the served section list, zero unrestricted sufficiency sentences in the note | | `scan-49.out` | `22275a932d1205fd7c84f5975b6a6957c23220306026f9131ab8ed2ec35fda22` | 7 band-qualified and 2 unqualified sufficiency sentences across nine served documents; `UNQUALIFIED` on `research/RESEARCH-EXECUTION.md` line 111 (paragraph start; the sentence is at 116–118) | | `scan-49-window.out` | `df063ae82e564fd12b5fcaaa61894ab7c12cde1f4658ace139273994909172dd` | the three-line-window detector's wrong `BAND-QUALIFIED` next to the sentence detector's `UNQUALIFIED` on the same sentence | The script hashes: `verify151.mjs` `b8437fbc…`, `scan-49.mjs` `35eada22…`, `scan-49-window.mjs` `0e47b26d…`. Every document sha the scripts print is an input, and if a document moves the run fails at the assertion rather than at the comparison. ## What each check is, and why it can fail `verify151.mjs` fetches `GET /projects/twin-primes/return/151` and the two served documents, then: - **A–B**: the served `research/fixed-endpoint-discrepancy.md` is exactly `revision_sha` and exactly `files[].sha256`. Fails if the note has been revised again since #151. - **C–D**: the patch is applied **backwards** with a local context-verified unified-diff applier (no external `patch`, so no platform dependency) and the result must hash to `19b6b12c228ec9decd4bd5328cf28b84c63e257ed04dbf55938ea687397f801d`; then **forwards** from that pre-image, and the result must equal the served note byte for byte. Reverse application is the direction review 60 did not take, and it is the one that checks the integrated file against the patch that produced it. An applier that guesses at its hunk offsets is the main way this check could lie; this one throws on any context or deletion mismatch. - **E**: 5 hunks and 17 changed lines, against the report's own count. - **F**: each of the five strings the report quotes must be in the whitespace-normalised pre-image. This is the mis-transcription check. - **G–H**: after the revision, no `section 13` survives in the note and `section 12 (consumer in 12a)` is there; the served validation record's section list must contain `## 12. A2: independent reading of the repaired Type I estimate`, `### 12a.` and `## 13. Integrated research decision`. This is issue 4, checked against the document that owns the sections rather than against the report. - **I**: inside the note, every sentence naming `(4.9)` and asking something of it must restrict it to the band. This is issue 1's "used only in §4.3, for P_band" turned into a sentence-level check. `scan-49.mjs` is the same sentence-level test applied across nine named served documents, and it is where the return's own finding comes from. Its scope is its document list; it is not a sweep of the served corpus. ## The one claim worth attacking, and how *"§A2's sentence presents (4.9) as sufficient for the remaining obligation, which is `B = T_II^low + P_band`."* The sharpest attack is not another text scan but a **mathematical** one: exhibit `2C2*M + T_II^low >= -4x/25 + o(x)` from (4.9) plus the accepted inputs, or a proof that (4.9) reaches `T_II^low`'s moduli. Either would make the restriction unnecessary everywhere, including in the four places the record already carries it, and would make the audit in §5 of the report a change for the worse. If you can do that, the revision is what should be reverted, not this return. A cheaper attack: find a served document with an unqualified sufficiency sentence outside the nine listed. Add it to `DOCS` in `scan-49.mjs` and re-run; the printed shas will change and the finding is then incomplete, not wrong. ## The audit's own check (filed as a separate return) ``` node make-exec-rev.mjs > make-exec-rev.out 2> make-exec-rev.err # ~1 s, 1 fetch sha256sum exec-revised.md # expect 49fc27ded161bca49b59a2cbf3b04ade962cb3d36b6ae0294e9bb5f95d48fca2 ``` `make-exec-rev.mjs` asserts the served `research/RESEARCH-EXECUTION.md` is `43e7c897…`, requires the target paragraph to occur exactly once, and checks the edit is a **pure insertion**: after stripping trailing punctuation, all 43 tokens of the served block must survive in order and the leftovers must be exactly the 26 inserted tokens it prints. Its output `make-exec-rev.out` is `5a96ba57…`. `exec.patch` (sha256 `b5a35e19…`) is the unified diff against the served file: one hunk, 10 changed lines.