# QC checks 9 and 10: `provenance` and `search-convention` **Date:** 2026-08-18. **Status:** built, calibrated, in the selftest, and deliberately OUT of the default gate behind `--pending`. Both were specified before they were built and both were deferred twice today for the same honest reason: they fire on live claims that an attack agent is mid-flight against, and that agent reports against a clean eight-check gate. This record is the price of that deferral — the full finding counts are here, so nothing is hidden. --- ## 1. What each check is, and where it was specified | check | specified in | catches | |---|---|---| | `provenance` | `research/PRIOR-ART.md`, "Provenance: what artifact was actually read" | a quotation of external work that never says **which artifact it was read from**; and a provenance note that names only a rendering, an abstract or a third paper | | `search-convention` | `research/SEARCH-CONVENTIONS.md` §6 | a claim that a **result is absent from the literature** whose paragraph neither cites `SEARCH-CONVENTIONS.md` nor names the convention it searched | `search-convention` is the check `absence` explicitly declines. `absence` drops any claim whose paragraph mentions the literature, because settling that means searching journals rather than listing a directory — so a claim that a *result* is absent was checked by nothing at all. That is the class that cost five audit waves. --- ## 2. Calibration, stated before any count ### `provenance` **Known positives: the five hand-read literature quotations in `research/two-class-lower-bounds.md`**, carried in `qc/lit-extract.js` and asserted before the check reports anything. If any is missed, the check reports **that and nothing else**, because a number from a sweep that cannot find what is known to be there reads as coverage. Two things were wrong with the calibration when this work started, and both were found by running it rather than by reading it. **(a) The list was keyed on file and line, and had already gone stale.** `lit-provenance.js` was written at 13:32 today and keyed its five positives to line numbers in `two-class-lower-bounds.md` as it then stood. Three commits the same afternoon — 13:48, 14:13, 15:17 — edited the document above those quotes. Every anchor drifted by 6 to 18 lines and the script began reporting `CALIBRATION FAILED` on five quotations the extractor was still finding perfectly well. **The keys are now content anchors**, a distinctive phrase from inside each quoted block, with the old line numbers kept only as comments. The framework's own rule for the `transfers` ledger says exactly this and said it first: key on content, never on position. The failure direction was safe — a stale key shouts rather than going quiet — but a calibration that cries wolf is one people learn to skip. **(b) With content keys, one positive was genuinely missed, and it was a real blind spot.** Positive #3 stopped being found because a correction pass earlier today rewrote `from their §1.3, verbatim:` as `from their **§1.1, p. 673** … verbatim.` — fixing a real citation error. The two asterisks between "their" and "§" broke the possessive cue, and the full stop after "verbatim" broke the verbatim cue. Both patterns were widened by exactly one character class each. **Effect on the inventory, measured rather than asserted:** 51 attributed quotations under the original regexes, **52** under the widened ones. The single newly reachable row is `research/two-class-lower-bounds.md:169` — which is known positive #3 itself. The widening was required by the calibration, not incidental to it. ### `search-convention` **Known positives: `research/two-class-lower-bounds.md` §2 rows 3 and 5 as they stood before today**, recovered with `git show 50f0307:...`: ``` | 3 | lower-bound construction for a 2-dimensional sieved set | **ABSENT** | see queries below | | 5 | the same accounting run in dimension 2 | **ABSENT**; done for the first time in §4 here | | ``` Both cells name no owning convention and cite no conventions file. **Both were false**, and both were corrected the same day: row 3 to `REFUTED — it exists` (Kalmynin–Konyagin, Izv. Math. 88:2 (2024) = arXiv:2302.00459 publish exactly that construction), row 5 to `NARROWED` (K–K do run a dimension-2 Mertens ledger; what survives is that *Maier–Pomerance's particular* ledger has not been run there). Verified by writing both rows back into a probe file and running the check: **both fire.** The two neighbouring rows that name `arXiv:1706.00317` and `A288815` — entries of the §1 table — **stay silent**, so the check discriminates rather than firing on the word `ABSENT`. **The negative control is the load-bearing half.** A paragraph reading *"nobody has published a bound on the fold, the tile or the zone at any level, and the scour was clean"* fires, and its note names the house terms it found: `tile, fold, scour, zone`. A calibrated negative in our own vocabulary does not satisfy the check, because that is precisely what five waves produced. --- ## 3. Live finding counts, 2026-08-18 Measured against the working tree at the time of writing. The corpus is being edited by a concurrent agent, so these will drift. ### `provenance` — **8 findings** Description line: *52 attributed quotations, 17 claiming verbatim, 10 carrying a note. 7 attributions are INHERITED from a source named further up and 3 resolve to nobody.* | kind | n | where | |---|---|---| | `unprovenanced-quotation` | 8 | `covering-dive.md:137`, `dhr-verification.md:191`, `dhr-verification.md:249`, `maier-matrix.md:154`, `maier-matrix.md:550`, `maxgap-law.md:363`, `two-class-lower-bounds.md:152`, `two-class-lower-bounds.md:190` | | `disqualified-provenance` | 0 | every one of the 8 quotations that *does* carry a note names a PDF, a page or a page photograph — none rests on an ar5iv rendering, an abstract or a third paper | | `extractor-uncalibrated` | 0 | calibration passes after the two repairs in §2 | `lit-provenance.js`'s own work-queue count is 9 rather than 8. The check drops one of them, `theta-ladder.md:52`, because its attribution resolves to nobody: it is one of our own custody warnings, not a literature quotation. All three `unresolved` rows today are ours, which is why that class is **counted in the description rather than reported as findings** — a kind that is 3-for-3 false positives fails the framework's own bar. ### `search-convention` — **75 findings** Description line: *89 literature-absence phrases, 11 name an owning convention or cite `SEARCH-CONVENTIONS.md`, 3 repeat a phrase already reported for their paragraph, 75 remain; 20 owning conventions and 7 house terms parsed from §1.* By document, heaviest first: | n | document | | n | document | |---|---|---|---|---| | 10 | `research/PRIOR-ART.md` | | 2 | `research/README.md` | | 9 | `research/covering-dive.md` | | 2 | `research/ZONE-POSTULATE.md` | | 8 | `research/sift-limit-attack.md` | | 2 | `research/natal-cap-04-packing-notes.md` | | 6 | `TODO.md` | | 1 | `paper/PAPERS.md` | | 6 | `research/G2-STATE.md` | | 1 | `paper/variance-note.md` | | 6 | `research/two-class-lower-bounds.md` | | 1 | `paper/writing-style-math.md` | | 4 | `web/PROPOSAL.md` | | 1 | `research/ATTACKS2.md` | | 3 | `paper/moire-primes.md` | | 1 | `research/NATAL-CAP-CAMPAIGN.md` | | 3 | `research/U-FRAME.md` | | 1 | `research/OBSERVATIONS.md` | | 2 | `paper/wall-note.md` | | 1 | `research/THE-DIALS.md` | | | | | 1 | `research/anchored-windows.md` | | | | | 1 | `research/gate-multiplies.md` | | | | | 1 | `research/h2-scoping.md` | | | | | 1 | `research/natal-cap-10-sieve-cap.md` | | | | | 1 | `research/two-moire-argument.md` | By trigger phrase: `nobody has` 17, `ABSENT` 13, `no published` 9, `unpublished` 7, `nothing published` 6, `not in OEIS` 6, `no paper` 4, `no prior art` 2, `possibly novel` 2, `not in print` 2, and one each of `no literature`, `not found in the literature`, `does not exist in the literature`, `apparently unworked`, `we found no`, `the first published`, `does not exist in print`. **Six `not in OEIS` findings are the highest-value ones in the list**, because that is the exact sentence the five waves got wrong: the `G₂` ladder has been in OEIS as **A144311** since September 2008, under wording containing none of our words. --- ## 4. The flip, and it is one line `research/qc.js` **line 73**, immediately above the `PENDING` map: ```js const PENDING_IN_GATE = false; // <-- THE FLIP. true promotes both into the gate. ``` `false` → `true` promotes both checks into the default run, into `--full`, and into `--strict`'s exit code. Do it when the running attack agent lands and its findings have been applied. Expect the gate to go red until the two queues above are worked. **Until then, nothing is hidden.** `--list` prints both under their own heading with the defect each catches. Every default run prints their names, why they are out, and a pointer to this file. `node research/qc.js --pending` runs them, and naming one explicitly (`node research/qc.js provenance`) runs it without the flag. --- ## 5. What could not be mechanised 1. **Whether an absence claim is TRUE.** The check asks only whether the paragraph says where it looked. `two-class-lower-bounds.md` §2 row 3 named no convention *and* was false; `research/README.md`'s index row names no convention and is merely describing what `PRIOR-ART.md` does. Only a reader separates those. This is the same permanent asymmetry `absence` records, one layer up. 2. **Whether a provenance note is TRUE.** `provenance` checks that a note exists and that it names an artifact of the right *kind*. Nothing checks that the PDF was actually opened. Requirement 5's rule applies here as much as to the ledgers: this can only be settled by a person, once, with a date. 3. **A one-line bold lead-in is treated as a list item** by the shared paragraph builder, so `**The Maier-Pomerance accounting, run in dimension 2 for the first time.**` is its own paragraph and its correction on the next line is not in scope. Here it happens to give the right verdict for the wrong reason — the joined paragraph would have cleared, because the correction names `arXiv:2302.00459`. Widening it would change `quotes` too, which is why it was left alone and written down instead. 4. **Known false-positive shapes in `search-convention`**, kept here so nobody re-fixes them by widening the suppressors: - an **index row** describing another document's job (`research/README.md:65`, "every claim as classical / obscure / possibly novel"); - a **style guide quoting a phrasing in order to ban it** — suppressed when the phrase is inside quotation marks, which handles `writing-style-math.md:122` but not the unquoted form at `:35`; - an **artifact absence sitting in a literature-shaped paragraph** (`TODO.md:255`, "nobody has tried bounding the residue-deleted maxsum"), which belongs to `absence` and already carries a dated `ABSENCE_VERIFIED` entry there. 5. **`SEARCH-CONVENTIONS.md` §6 is now stale in one word.** It says "Nothing mechanical enforces §1 yet". Something does, but it is not in the gate, so the sentence is still true of the gate. Left unedited on purpose: that document is inside the blast radius of the running attack agent, and the honest rewrite is one line at the same moment as the flip. --- ## 6. Suppressions added **None.** `qc/ledgers.js` is unchanged. Neither check needed an entry, and the only key either of them uses is content: `search-convention`'s one-finding-per-paragraph dedupe is keyed on the paragraph's own normalised text, so an edited paragraph returns rather than staying suppressed. --- ## 7. Files touched | file | change | |---|---| | `research/qc/checks.js` | added `provenance` and `searchConvention`; `quotes` now calls the shared paragraph builder instead of its own inline copy (verified output-identical) | | `research/qc/corpus.js` | added `paragraphIndex()` (lifted from `quotes`) and `SEARCH_CONVENTIONS` config | | `research/qc/lit-extract.js` | **new.** The one extractor for attributed quotations, moved out of `research/lit-provenance.js` | | `research/lit-provenance.js` | now the report over `qc/lit-extract.js`; calibration keyed on content; two cue patterns widened | | `research/qc/selftest.js` | copies `lit-extract.js` into the fixture; 6 new positives, 5 new controls | | `research/qc.js` | `PENDING` map, `--pending` flag, `PENDING_IN_GATE` flip, visibility notice on every default run, `--list` section | | `research/qc/README.md` | documents both checks, the flip, and the `sourcing` row that was missing from the table | **Why the extractor moved to `qc/`.** `checks.js` first required `../lit-provenance`, and `selftest.js` builds its fixture from a fixed copy list, so the import did not resolve inside the fixture and the whole selftest died. The fix is *not* a tolerant `existsSync` probe: an extractor silently resolving to nothing makes `provenance` report zero and read as clean while checking nothing — the exact failure `ledgers.js` was moved off this morning. The shared code now lives under `qc/`, both callers require it hard, and the fixture copies it.