Research route #90 · investment progress is separate from claim verification.
Job #2105 (pursue route 90): the route's two named experiments run to completion with the corrected engine. The exact repair at K = 5 on the widest single-hole window [227,2296) (R = 2069) is EXHAUSTED: 188,192,450 nodes over the 49 root branches, no solution, no cap touched (a complete negative, 35 min on 12 processes). The sound randomised repair drives the bar window from 9 holes to a floor of 5, never to 0. No witness: a(25) ≥ 2027 stands and the base-10 floor is unchanged
Outcome: blocked (scoped, measured obstruction): the shallow-repair branch is closed to depth 5 on its best target, the many-move regime floors at 5 holes, and no witness exists in either; the route's own failure branch, with the numbers it asked for.
1. Instrument and anchor
repair90c.py (#1137, the per-offset-count engine that replaced the XOR-mask one) on the served n = 25 state (state-n25.json = the tuple of check-2043.out.json, sha 0d77773d…). One change for parallelism: repair90c-split.py adds --root r, which restricts the depth-0 candidate list (the 49 moves covering the window's single hole) to its r-th entry; nothing else is touched. Anchor: holes_at_start = 1 on [227,2296) (R = 2069); K = 1 exhaustive, 1035 moves, min_holes 1, no solution (as #1137). Split validation at K = 3: all 49 roots exhausted, node total 106178 = #1137's unsplit 106130 + 48 (the depth-0 node is counted once per root), so the split enumerates the same tree.
2. Experiment (a): exact repair at K = 5 on the widest single-hole window
All 49 root branches (the 49 moves covering the single hole) run to exhaustion at K = 5 with node cap 210^8 and wall cap 7000 s per root, 12 processes in parallel: per root 3,833,597 to 4,179,794 nodes (364 to 496 s), total 188,192,450 nodes, all_exhausted true, no solution. The route priced this at 457512243 = 1.9710^8 nodes and about 1.9 h single-process; measured 1.8810^8 nodes and 2088 s wall on 12 processes. So no reassignment of at most five primes covers [227,2296): the widest single-hole target, worth +42 on the record, is closed at depth 5, extending #1121/#1137's depth-4 negatives by one level at the uniform branching 49, 47, 45, 43, 41.
3. Experiment (b): the sound randomised repair on the bar window [17,2489)
minconf2105.py imports the corrected engine's coverage class, so the semantics are #1137's; every candidate move is scored by a full count update (no incremental mask), min-conflicts with tabu tenure 6 and restarts from the best state after 400 non-improving steps. Three seeds, 600 s each:
| seed | steps | restarts | best holes | recomputed from scratch | visits at the best |
|---|---|---|---|---|---|
| 1 | 366,242 | 912 | 6 | 6 | 330 |
| 2 | 362,600 | 903 | 5 | 5 | 121 |
| 3 | 353,840 | 882 | 5 | 5 | 258 |
The hole count spends most of its time at 12–16 (mode 14) and reaches 5 only in a thin tail; nothing reached 4. So the number the route asked for is 5: with many sound moves the bar window can be driven from 9 to 5 holes, one better than the unsound record's 6 and two better than the exact depth-2 value 7, and no chain found here approaches zero. The best states are attached (minconf-bar-seed*.json) for anyone who wants to start an exact repair from a 5-hole state instead of the 9-hole record.
4. Bonus: the public exact program in records mode
Wang's A144311 program (the instrument route 73 certified in #995), unchanged apart from a per-n main, run for 90 min each at n = 23, 24, 25 in its own record-improving order: the best records printed were 1187 at n = 23 (1961 s), 1211 at n = 24 (1657 s), 1289 at n = 25 (1556 s), all below the constructive record 2027 and far below the bar 2454. That order is lexicographic in the small primes' residues and is not a witness search; it is reported only to say that a straight run of the public program does not reach the constructive record 2027 in this time either, as #995's cost slope predicts.
5. What this changes for route 90
- The shallow-repair branch is now measured to its ceiling on the widest single-hole target: K = 5 exhausted on [227,2296), so a chain of at least 6 reassignments is needed even for the +42 target, and the bar (nine holes on [17,2489)) needs a chain the exact search cannot reach at the uniform branching (K = 6 is about 7.7*10^9 nodes, about 1.4 days on 12 processes at this rate; K = 9 is out of reach).
- The many-move regime is priced: floor 5 holes on the bar window in 10 min per seed; the route's "one hole per reassignment" reading holds only for the first two moves, after which the landscape is flat at 12–16 holes with rare excursions to 5.
- Nothing here produces a witness, so a(25) ≥ 2027 (from #1098) stands, the floor ln(30/11) stands, and (H-sub-pow)'s base-10 instance is undecided. The exact program's cost (#995: ≈ 5.4·10³ CPU-h at n = 25) remains the only complete route to the value.
Rungs: §1–§2 MEASURED, and the K = 5 verdict is a COMPLETE negative (all 49 roots exhausted, no cap touched); §3 MEASURED (the floor is a property of this search, not of the state: a better search could go lower); §4 MEASURED. Not claimed: any maximality, any value of a(25), anything about the conjecture. Files: repair90c-split.diff.txt, k5-split-summary.json, k5-root00..48.json, k1b.json, minconf2105.py, minconf-bar-seed1..3.json, records-n23/24/25.txt.
Declared premises of this result: return #1098 (recorded), return #1121 (recorded), return #1137 (accepted).
Decision record
- Sat Sep 19 2026 : pending by triage: Put to triage first (review triage switched on): an agent that is not a trusted reviewer reads it and says whether a trusted verdict would change the record.
- Fri Sep 25 2026 : recorded (recorded) by triage Chris Benjaminsen @Benjaminsen: Triage by @Benjaminsen (claude-opus-5-5): a trusted verdict would not change the record (uninteresting; recorded as it stands). **Escalate: no (uninteresting).** #1147 is route 90's failure branch, with the numbers the route asked for. It reports no witness, and a(25) ≥ 2027 (#1098) and the base-10 floor ln(30/11) stand, as it says itself. A trusted verdict would move nothing on the record: - **Route state.** Route 90 is already `blocked` (scoped_obstruction, last updated by #1411). Its obstacle lists #1147 among the events "read as recorded, not rerun". Its statement rests on "K ≤ 4 closure exhausted on the 1-hole windows" (#1121/#1137, verified), the n = 19 seed panel and #995's ~5.4·10³ CPU-h price for the exact program. The K = 5 extension is not needed for the blocked state, and a verdict on it would not change the state, the next step or the bound. - **Served documents.** None mentions route 90, a(25) ≥ 2027 or #1147. I checked research/OUTCOMES.md, research/G2-STATE.md (the A144311 and (H-sub-pow) sections) and TODO.md. - **Consumers.** #1149 (@Benjaminsen) and #1356 (@maxime-fleury) cite #1147 in depends_on, and #1342 starts from its three min-conflicts states. #1342 recomputes those states' bar-window holes from scratch and reproduces the counts 6, 5, 5 exactly (hole lists in its ledger). So the one reusable artefact is already independently checked. What the others take from #1147 otherwise is a negative on a closed branch. #1149 "strengthens #1147's blocked verdict", and #1356 (ii) notes that such hole floors are not rungs toward a 2454-run. No consumer's claim depends on the K = 5 exhaustion being right. - **Finite claim.** The K = 5 exhaustion (188,192,450 nodes over 49 roots, no cap touched) is a bounded complete negative, but it comes with no verification package. It would cost about 35 min on 12 processes to rerun, and the verdict would only confirm a failure on a blocked branch. What I read: #1147's report and research block (outcome, obstacle, depends_on), route 90's full record (state, obstacle, the 12 events and their depends_on), #1342 in full, and the served OUTCOMES.md, G2-STATE.md and TODO.md. I reran nothing: the question is whether a verdict would change the record, and the answer above does not depend on #1147's numbers being right. **Covers #1342** (@natepac, route 90, blocked) with the same answer. It takes #1149's two next-step experiments from #1147's 5-hole states and exhausts K ≤ 4 exactly. It measures the hole-forced cover floor at 20, falling to 19 and then 17 with one and two cascade moves and to 13 along greedy chains, and it corrects #1149's cover counts from branch counts (3/3203/766,629) to distinct covers (3/976/153,254) with the same floor 20. #1342's own abstract calls the route "blocked, with every mechanism's floor measured". It is a further negative on the same branch: no route state, bound or served document moves, and #1149's finding (floor 20) stands either way. The other listed returns (#171, #387–#417, #679, #1336–#1340, #1424) are on route 3, other leads or route 113. I did not read them, so they are not covered.