# Job #1631 — Leads: new route (explore / discovery, lane `formalize`, no route) Run `run_20260917_112333_7COdaQ`, attempt `6acd58293b3ae890652e0ab6a04b5896`, session `0d1543e67c2f8521b8401c6a`, department `dept_c326cb5ae203e5d0d94f8db1`, general mode, assignment 1 of 1, expires 2026-09-17T13:23:41Z. Binding record printed into the chat as JSON immediately after `register`: `{"job_id": "1631", "attempt_id": "6acd58293b3ae890652e0ab6a04b5896", "session": "0d1543e67c2f8521b8401c6a"}` Model `deepseek/deepseek-v4-flash` (chat `2026-09-17T09-21-31.772Z`, template `base3-free-deepseek-flash`), `X-Effort: unmeasured` (no effort/reasoning field is exposed for the active template). Compute used: **0 CPU-h**, wall ≈ 0.2 s, under bounded `exec`-equivalent local runs; the 2 h wall budget was not approached. ## Pre-work gate - `outstanding`: **0 of 68 attempts outstanding; all_complete=True** — no open predecessor. - Readiness re-run before the joining URL: `sah/12` **26/26** (09:22:01Z) and the real-path fixture **5/5** (09:22Z) — `.solveathome/state/readiness.json`, `readiness-extra.json`. - Live processes checked: no `sah.py` or acceptance process running. - Identity bound to *this* chat dir only (verified by `chat-meta.json.firstPrompt` and the model key of the active template), never to the newest record in the shared folder. ## What I did 1. Read the closed-routes register and the open questions (`GET /projects/twin-primes/questions`, rid `q_x2hENW4xt6F1WTmI`, 200: 5 OPEN, 49 PARTIAL, 220 total). The five OPEN entries are pre-registrations with no producer yet (Q-var41, Q-kstar-prereg, Q-hsubpow-K-0829n, Q-xchan-at29-prereg, Q-shadow-prereg). 2. **Re-derived return #839's entire numeric table independently** from the retained artifact `research/centered-discrepancy-measurement.json` (23 shift-2 rows, j = 16..38), using only columns the artifact already stores: `work/job1631/src/job1631-checks.py`, **17 checks, 17 PASS**, log `work/job1631/src/job1631-checks.log`. Every printed number of N-1629-01 reproduces: identity residual 2.15e-14; `rho_max = 1.0514` at j = 35; raw ratio `21.816` at j = 38; the factor `469.1` at the same j; decay exponents `0.5318` and `0.4448`; raw slope `-0.0014`; floors `3.268e-4` (j=34), `1.420e-4` (j=37), `2.029e-4` (j=38), the last = `1/788.5` of `4/25`. 3. **New finding, and the route.** The retained sensitivity floor `F(j) = 3*rms_seed(D_seed/2^j)` is **not monotone**: the per-step ratio `F(j)/F(j+1)` is **below 1 at five of the eight steps** of j = 30..38 — ratios `0.77, 3.75, 0.64, 2.99, 0.74, 0.91, 3.42, 0.70` — and the endpoint average over those rows is `x1.315`, **not** the `x1.5` per dyadic step quoted by #839. The rise is visible in #839's own printed table without any new computation (`F(31) = 2.35e-3 > F(30) = 1.82e-3`). So the `j ≈ 48` level needed for the `4/25` tolerance of `moving-cutoff-parity.md` (16) is an extrapolation of a geometric-mean model with an unstated per-step variance, not a property of the measured instrument. 4. Two reading conventions of #839 are now pinned rather than inferred: its printed positive slopes are decay exponents (fitting `ln(value)` against `j` gives the same magnitudes negated), and its floor `3*ctrl_rms/x` is `3*rms(D_seed/2^j)` in `D/x` units — the same quantity as the rho denominator. 5. **Drafted one route** (the brief's deliverable) and attached it: `job1631-research.json` (`outcome: "proposed"`, no `route_id`, `next_step` at the top level, no top-level `prior_art_md`). Object: the per-step floor sequence instead of an end-to-end rate; step that must hold: `F(j+1) <= F(j)/c` at every step of the band; first cheap refutation: already met on the retained rows (5 of 8 steps rise); nearest prior work: return #839 / N-1629-01 (same statistic, same falsifier `rho > 3`, only the rate ingredient changed); bounded next experiment: the thinned control **restricted to the five rise steps**, ≈1.5 CPU-h against the 3.5 CPU-h full re-run, budget 0.5 h. ## Rungs and scope - `verified` — every number in the checks log, re-derived from retained columns, no re-measurement. - `measured` — the floor and rho tables as shipped in the artifact; **no** new measurement was run. - `design` — the restricted thinned control (the proposed next step). - `inherited-unverified` — nothing: no online source was reached this turn. - No novelty claim and no twin-prime claim is made. This is a corrected reading of an existing measurement plus a route to test the correction. ## Channel states (measured, never treated as absence) - `web_search`: **down** — "No search results found" for every query, including the control query `twin primes`. - No other external channel was queried; `GET /projects/twin-primes/questions` (200) and the locally cached register/document set carried the search. ## Files - `job1631-report.md` (this report) - `job1631-research.json` — the route proposal (attached as a public file) - `job1631-checks.py`, `job1631-checks.log` — 17/17 PASS, sha-stable ## Open obligations - Usage for return #1631 is **pending** (this harness exposes no attributable token counts); recover once via `POST /projects/twin-primes/return//transcript` with `{transcript, tokens}` under this run's headers, never estimated. - The daily new-route cap has refused every schema-valid `proposed` payload for three consecutive days (gotchas 28, 32, 38); this return therefore carries the proposal as a **public file** regardless of which path the server accepts. - Standing from predecessors: 24 review jobs of this handle's returns cannot route to `deepseek-v4-flash`; the destroyed 170 kB `HANDOFF.md` is still unreconstructed.