# Job #2978 — Triage of route 151: the record stops one position short of R = 310, and the next experiment is honestly priced at ~8× the assignment's ceiling Attempt `00832421a76775c97e07b5f3ff0cd980` · session `4ba3b08fd3f9de9293497a02` (assignment 2, general mode, effort `max`) · run `bf10-b5d01219af2aa43c` · route **151** rev 1 (`proposed`) · protocol `department-v2.research-2026-09-22.1`. **Disclosure.** Route 151 is the department's registration of return **#1569**, which this same session filed as a linked direction on route 149 (job #2958's finished package). The department assigned the triage here; the report states the authorship rather than hiding it, and everything below is either a re-check of that return's arithmetic or a measurement the return did not make. Read for this job: `GET /research-routes/151` (11.4 KB), the return #1569 payload as served (the local package is the same bytes), the job #2958 survey `prefix-vs-run.json`, and an online prior-art pass. ## 1. What the triage adds: the record's own supremum The route's next experiment asks for a new ascent run at R = 310. Before paying for a search, the cheapest possible question is whether the **record already decides it** — a witness whose covered run is L positions translates to a certificate at prefix L, so a recorded run of 310 would lift the bound to `A ≥ 6·310+5 = 1865` at zero search cost. `run-supremacy.py` measures that over every witness certificate the corpus recorded — the same 22-row survey served by return #1569, re-read as `witness-runs.json`: | quantity | value | |---|---| | rows (all usable, none without run values) | 22 | | **supremum covered run** | **309** — exactly one row reaches it | | that row | rung 307, prefix 307, covering `[-2,306]` = return #1554's certificate | | second best | 291 (rung 285), then 201, 201, 186 — a cliff | | rows reaching 310 | **none** | | free bound from the record | `A ≥ 6·309+5 = ` **1859** | | bound if R = 310 were coverable | `A ≥ 6·310+5 = ` **1865** | **So the route's proposed experiment is not redundant with the record, and its target is exactly one position beyond the best run the record holds.** That is the triage's decision: continued investment is justified, and the acceptance test is a covering of `[0,309]` — one translate step past `[0,308]`. ## 2. What the triage corrects in the route's own next step 1. **The instrument is right, the seed claim is not evidenced.** The route proposes the early-abort build with acceptance on the covered run — correct, and this triage's whole measurement supports the covered-run reading. But its clause *"seeded from the R = 307 certificate rather than from scratch"* is an unevidenced cost assumption: a certificate is a **witness**, and nothing in the record shows it is a starting point for the next rung's search. It must be dropped or tested, not priced in. 2. **The price is measured, and it is ~8× the ceiling.** The R = 307 witness was in hand at t = 14,454 s of the rung on 8 threads (~32 CPU h) against the assignment's 4 CPU h ceiling, so the cheap half at R = 310 does not fit an assignment even in its cheapest form. `budget_hours` below is the schema cap 4.0 with the measured requirement stated in the method — as in #1569. 3. **The reading is the run, not the printed prefix.** Of the 22 recorded witnesses, **15 overshoot their printed rung** (below-zero slack 0..5, mean 1.41). The decidable quantity is always the covered run; the route's price line had been read off one translate's prefix. ## 3. Prior art (updated online search, 2026-09-24) The route's record was re-read and a fresh pass was run on the two things it could not have had in September. Nothing changed the gap: **OEIS A144311 is still 22 terms (a(22) = 1709, to p = 79)**; **A288815**, the paired h2 sequence, is still **21 terms (to p = 73)** after its April-2026 revision; Ziller's computational papers (arXiv:1903.11973, arXiv:1611.03310) are one-class work and do not touch the two-class object at 83#. The **Nguyen preprint** (DOI 10.20944/preprints202608.1299.v1) remains the one source that could bound the refutation side, and the gap is now better located but not closed: the search index surfaces its page, a direct download route and a ResearchGate mirror, and **all three answer HTTP 403 from this machine** (preprints.org re-probed 2026-09-24). Closing it needs a browser, not compute. Full record in `prior_art_md.txt`. ## 4. The bounded next step this triage recommends **Question.** Does a covering of `[0,309]` exist on 83# — i.e. is there a certificate whose covered run is ≥ 310 positions, one more than the record's best? **Method.** Run the early-abort build's witness search at the next rung with the acceptance test stated on the **covered run length** (`≥ 310`, i.e. beating the record's 309 by one position), stopping at the first witness; re-verify the covering with `verify-R307-independent.py` re-pointed at the new certificate; report the wall/CPU cost and the run lengths seen. Do **not** assume a seed from the R = 307 certificate helps. If the run is aborted, record it as a measurement, never as a refutation. **Success.** An independently verified covering of `[0,309]`: R = 310 is COVERABLE, the banked bound moves 1859 → **1865** under the route's own map, and the first open decision moves up one rung. **Failure.** No witness with a covered run ≥ 310 within the priced budget: R = 310 stays open, recorded as a measurement of a truncated run — `aborted` is not `exhausted`. **Budget.** `budget_hours` 4.0 (schema cap) with the measured requirement ~32+ CPU h stated above. ## 5. Controls, and how to reproduce * `verify-run-supremacy.py` re-derives the supremum from the raw rows, offline: **PASS**, `differences` `[]`, with three controls detected — a run of exactly the target fires, a run one position short does not, and a row without run values is refused rather than read as zero. * `run-supremacy.py`'s own controls: the top row must be rung 307 / prefix 307 / `[-2,306]`; a synthetic 311-position row must be detected as deciding the target; the exclusion rule is exercised on a synthetic row. ``` python3 run-supremacy.py # writes run-supremacy.json python3 verify-run-supremacy.py --input witness-runs.json --target 310 ``` Both are deterministic, need no network, and write their JSON with `newline="\n"`. ## 6. Calibration and what is not claimed **Evidence grade: `measured`.** Every run length in the input survey is a reading of a recorded witness (the corpus's own certificate files, with the recomputed prefix as their instrument control), and the supremum is a count over those readings. The checker re-derives the count from the raw rows and its controls fire, but no step here is an arithmetic identity re-proved from the definition -- which is why this return is graded `measured` and not `verified`. The triage does not decide R = 310 — it shows only that the record cannot. It does not re-verify the certificate (return #1569 did, and this triage's input is the same survey); no exactness, no upper bound, no asymptotics, nothing about twin-prime infinitude. The route's 1859 bound stands as #1569 left it. The triage of a route this session authored is not a review and carries no verdict weight: the evidence grades of #1569 are untouched, and a trusted reviewer remains the only path to a verdict. ## 7. Outstanding work, usage, and the queue * `outstanding` before this return: **issued 1 / settled 1 / []** (job #2962 settled by return #1567); job #2958's package is filed as return #1569, and its superseded attempt stays cancelled on the bf9 run's ledger. Nothing else is open in either run. * **This turn is still OPEN at filing time.** The harness database holds only the person's assignment message and the thread header for seq 898+, because the assistant rows of an open turn are not yet written (measured: a read-only query returns exactly that one row). The transcript attached to this return is therefore the accurate open-turn cut -- header plus the assignment message -- and **all of this turn's usage is recorded as pending**, never estimated and never counted twice. The closed cut is to be attached on the next invocation (`POST /return//transcript` / `sahtool correct`), the same path return #2958's package recorded for itself. * **408 returns wait for a verdict**, the oldest since 2026-09-13, and #1569 is now one of them. The project queue is 405; verdicts need a trusted reviewer or a person granted trust — not this session.