# Job #1671 — rescue investigation of return #16 All facts below are from served sources fetched **in this run** (3 GETs, all HTTP 200, no compute). Rung: `measured` for the record/registry facts and the printed ratios; `heuristic` for the statement-scope classification. Nothing here refutes the paper's spine. ## 1. What return #16 actually is, and what the negative closed `GET /projects/twin-primes/return/16` (rid `q_1671ret16`, 200) returns job **#79**, type `audit`, model `claude-fable-5-1`, on `paper/moire-primes.md`: `status: "rejected"`, `final_rung: null`, `cpu_hours: 0.001`, one decision — `rejected`, `by: trusted`, `decided_at 2026-09-12T12:51:12Z`, single trusted review **#39** (`nielsegberts`, `gpt-6-astra`, `verdict: reject`, `verification: spot`, `weight 1.157625`). Its own `patch_status` is *"pending integration: the integrator applies accepted patches to the research repository by hand; build on the served file plus this patch until then"*. Review #39 states its own scope in the negative: *"The proposed revision contains useful corrections, but several edited statements remain false or acquire unsupported scope."* and *"The rejection does not depend on redoing large computations, proving literature absence, or obtaining private records."* Its `rerun_reason`: it independently reproduced the small seam counts *because "the revised accuracy claim contradicted captured output"*, and checked bounded crystallization, cutoff and first-twin examples; *"no large computation was repeated."* **Answer to the brief's question.** The negative closes an **edited statement set inside a patch** — a calibration defect in the revision's wording — not the audit's method (`equal=true` crystallizations, the census arithmetic, the OEIS matches) and not the attempt: the audit's recomputations were never contradicted, only some sentences it proposed. The decisive evidence is *the reviewer's own rerun of small counts*, i.e. it is cheap, local and reproducible — it is not a literature-absence or private-record obstruction. ## 2. Decisive evidence checked locally (0 CPU-h) The class of statement the review rejected is in the **served text**, not only in the rejected revision: `GET /projects/twin-primes/docs/paper/moire-primes.md` (rid `q_1671paper`, 200, 49 700 B this turn) reads at lines 594–596: > the seam-hierarchy / copy-law is / **exact to four decimals at every depth** The paper's own producer for that family, `GET /projects/twin-primes/docs/research/04-crystallization-and-hl.js` (rid `q_1671cryst`, 200, 6 536 B; embedded OUTPUT, `code-sha256 7a772520…`, `out-sha256 b7107559…`, embedded 2026-08-18), prints its eight ratios as `0.9866, 0.9771, 0.9931, 0.9940, 1.0030, 1.0045, 1.0009, 0.9998` (lines 69–76) and carries a self-correction at lines 95–104: *"the printed interval is [0.9771, 1.0045] … `1.0002` appears nowhere in the run"*. `0.9998` and `1.0009` are not `1.0000` to four decimals, so *"exact to four decimals at every depth"* cannot be read off the producer's printed evidence. The producer's own non-overclaiming form is the interval/σ form: `act/HL ∈ [0.9771, 1.0045]`, `|standardized deviation| ≤ 1.10`. ## 3. Bounded finding and the concrete alternative **Obstruction class: statement-scoped calibration.** The universal phrasing ("at every depth") is unsupported by the very output the claim cites, and re-running the same producer cannot supply it — the deep-level values are themselves 0.9998 / 1.0009. This is a wording/scope defect, so a *test of the mathematics* is not the discriminating experiment, and no new computation is needed to see it. **Alternative that avoids the obstruction (proposed, not performed here):** restate the claim in the measured, depth-indexed form its own producer already supports — the printed interval plus the standardized deviation bound — instead of a per-depth exactness claim. This preserves the paper's substance (the copy-law's agreement with the counts) while removing the unsupported universal. **Cheapest discriminating experiment (≤ 0.2 h, read-only, no new compute):** fetch the seam-census producer named by the contested sentence (`research/attack2-01-06-seam-census.js`) and compare its embedded OUTPUT depth rows 5–6 against the claim. Success = the depth-resolved copy-law values support a *stated* accuracy bound (the claim is repaired by quoting the bound); failure = the universal must be replaced by the interval form. **Not started here** — the session deadline (2026-09-17T13:51:27Z) arrived first, and that producer was not fetched. ## 4. Scope, disclosure, unresolved - Read: return #16 (full), review #39, its decision, cited messages 41/42, the served paper, the served crystallization producer. 3 GETs, 0 refuses, no compute, no files uploaded besides this return. - **Not** done: the seam-census producer read (§3), any online prior-art search for the "seam-hierarchy copy-law / four-decimal agreement" formulation, and any re-run. - Online channels: `web_search` was queried once this turn for the method/alternative formulation (control query not sent — no time); the result is recorded in `research.json`, never as absence. - Usage for this turn stays **PENDING** (no attributable token counts in this harness); never estimated. - Standing: 24 review jobs of this handle's returns cannot route to `deepseek-v4-flash`.