# Pre-registration — job #2972, route 114 (audit-revision freshness) Written 2026-09-24T01:5xZ, BEFORE any new network fetch by this job. Frozen text; any later change is an appended amendment with its own timestamp, never an edit. ## What is already measured (CITED, not re-derived here) * #1373 (population): ids 1..1370 walked, 1200 pages found; status census over those pages accepted 101 / recorded 626 / pending 424 / rejected 47 / superseded 2; of the 101 accepted, 36 carry a patch or a revision, split **11 audits, 11 break, 11 measure, 4 explore**; 59 candidate (return, target) rows. Enumeration artifact sha256 ed06d867…. T5 in that return: for #101 / #151 / #152, removed lines still served 5/5, 7/7, 1/1 and added lines present 0/25, 0/9, 0/1. * #1358 record census (as quoted by the route): audit accepted 2, measure accepted 24, break accepted 17 — inconsistent with #1373's 11/11/11/4 and never reconciled. * #1354 / verdict-drift-check/1.0 (sha ab5905ac…): 4 of 4 audited documents still serve the pre-audit text, none serves the revision, 3 of 3 index rows publish the pre-audit verdict. Population = the four rehearsed audits #101, #151, #152, #153. * #1447: /docs listing at 2026-09-22T22:41:53Z, 1211 paths, listing sha256 46ac650e…. * Route 114's own unresolved step 3, verbatim: *"any accepted audit whose revision IS in the served text" is used below as the falsifier, and it is a search over an unknown-size lane; if the audit lane contains only these four returns the search is trivial, and if it contains many, the search is where the hour goes rather than the conclusion.* ## Question this job runs on Two questions, one instrument, both decidable from served bytes read anonymously: * **Q1 (lane size).** How many returns of `type = audit` exist at all, how many are accepted, how many of the accepted carry a patch or a revision (the only two carriers of a *stated* revised verdict), and is that set the four rehearsed audits or larger? * **Q2 (counterexample search, lane-wide).** For every accepted audit that states a revision, is that revision served — (a) in the document the patch's own header names, (b) anywhere in the 1211-path served corpus, (c) in the ledger layer the corpus ships (`research/qc/ledgers.js`, `research/QUESTIONS.md`, `research/history/staging/audit-ledger-extraction.md`)? Any (a) or (c) hit is a propagation path and refutes the route's premise. ## Population definition (fixed here, before fetching) * A = returns with `type == "audit"` found by an anonymous GET of `/projects/twin-primes/return/` for ids 1..MAX, MAX chosen as the first id above the highest id that answers 200. Anonymous means no Authorization header at all. * A_acc = a in A with `status` in {"accepted","proven","verified"} (recorded as served). * A_rev = a in A_acc with a non-empty `patch` or a non-null `revision_sha` — the audits that state a revised verdict. **This is the set the Q2 search must cover**; the four rehearsed audits are a subset of it or the instrument has failed. * Target resolution: the path in the patch's own `+++ b/…` header, falling back to `--- a/…`, then to the return's declared target. (The declared target is a candidate pair, not a fact: #1373 measured 18 of 59 rows as WRONG-TARGET.) * Line test (literal, no fuzzing): for every line of the patch beginning `+` (not `+++`) strip the marker and require it verbatim in the served bytes; likewise for `-` lines (not `---`). Empty/whitespace-only lines are excluded from both counts and reported separately. ## Predictions (each with the falsifier that kills it) * **P1** |A| >= 8: the audit lane is larger than the four rehearsed audits, and |A_rev| >= 4. *F1*: |A| <= 4, or A_rev is exactly {101,151,152,153}. Then Q1's answer is "the search was already lane-wide" and this job records that and stops rather than manufacturing work. * **P2 (control, must pass for Q2 to mean anything)** The four rehearsed audits are in A_acc, and my line test reproduces #1373 T5's numbers for #101 / #152 / #151 up to one line. *F2*: any reproduction 0/0 or off by more than one line. Then my patch reader, not the finding, is what differs, and Q2 is void until fixed and disclosed. * **P3** No accepted audit in A_rev has any of its added lines served in the target its own header names: counterexamples = 0. *F3*: >= 1 audit whose added lines are served verbatim in its own target. Then a text-layer propagation path exists and the route's premise fails. * **P4 (the central falsifier)** No accepted audit in A_rev has its revised verdict carried by the corpus's ledger layer: for every a in A_rev, the served target's ledger block and `research/QUESTIONS.md` do not carry a's verdict text (the phrase its patch adds) and do not record a's own return id as the source of a verdict differing from the served one. *F4*: any a in A_rev whose verdict text or id-with-revised-verdict is served in that layer. Then propagation happens at the ledger layer, the corpus is *not* silent, and the correct repair moves from "add a carrier" to "make the two carriers agree" — a different finding. * **P5** For at least one audit row the patch's header target differs from the return's declared target (#1373's WRONG-TARGET effect reproduces inside the audit lane). *F5*: 0 disagreements, in which case the declared target is reliable for this sub-population and says so. * **P6** 0 of the Q2 line tests depend on a fetch that needed a credential: the whole measurement is anonymous. *F6*: any 401/403 on a page used for a number. ## Controls * **C0 (corpus unmoved)** Re-derive the live recursive /docs listing in this job and compare it with #1447's frozen listing (1211 paths, sha256 46ac650e…). If it differs, every comparison with a cited corpus number is re-scoped to "as of this job's listing", and the new listing sha is reported. * **C1 (negative control, feed-tamper)** Deliberately test one audit against a document it does NOT name, and require zero added lines present. A test that reports "present" for a mismatched pair is measuring something other than presence. * **C2 (positive control, instrument can see text it should see)** For one audit, require its *removed* lines (which are the pre-audit text, thus expected to be served) to be present at a rate strictly above the added-line rate. A test that cannot find *anything* is broken, not a finding. * **C3 (id-walk completeness)** The walk reports 404 count and the highest 200 id, so the lane census is bounded by measurement, not by assumption. ## Decision rule, fixed before the run * F1 fires → outcome `known` (or `inconclusive` if the census contradicts itself): record the lane size and stop; no next_step. * F3 or F4 fires → outcome `result`: a propagation path is named, the route's central question is answered, and the next step is to make the two carriers agree, not to add one. * No falsifier fires → outcome `progress`: a lane-wide census plus a re-runnable `audit-gate.py` reporting exactly the audits whose accepted verdict has no observable product; next_step = run the gate after the maintainer action, plus the cheapest repair (record the carrier, or hash the patched text). Compute stays read-only and anonymous. ## Scope, stated in advance Served endpoints only (`/return/`, `/docs/`, `/history/`, `/files/`), anonymous wherever a claim is about a reader. The private repository is not read, no audit's mathematics is adjudicated, no note is edited, and `patch_status` values are treated as claims about integration, not as evidence of it. Nothing here is claimed about returns outside the walked id range.