A mid-game depth-7 decision cost more than three times the opening

Yesterday's depth-7 work ended with a measured opening decision — 2302 s at 0.983 worker occupancy — and a two-hour per-decision budget sized with what looked like threefold headroom. Playing the same policy through gauntlet-01 disproved that sizing at move 9.

negative

The depth-7 opening decision is not representative of the game

mid-game decision exceeded 7200 s, opening cost 2302 s

rust-fair-d7-s7 completed eight moves of gauntlet-01 and was then killed by its own budget on the ninth: board 0000000000000000000000002000000670084989098888898 (three occupied rows and part of a fourth, several covered discs, most columns still open) did not answer within 7200 s under the frontier scheduler at D7/S7 with a 16,777,216-entry cache. That is at least 3.1x the measured opening cell, and it is a lower bound only — the process was terminated, so the position's true cost remains unknown. The opening position is therefore a poor basis for estimating a full depth-7 game, and yesterday's "order of a day and a half" figure should be read as a floor rather than an estimate. Record: runs/RUN-20260828T0700Z-d7-midgame-timeout/.

The budget was raised to four hours. This is an operational guess, not a measurement: nothing has yet established what the expensive mid-game positions actually cost at depth 7, only that at least one exceeds two hours. Whether four hours is enough for this position, let alone for the fuller boards that follow it, is open.

positive

Crash-safety held: the overrun cost one decision, not eight

8 of 8 completed moves retained

The overrun surfaced as the harness's own budget error rather than a raw spawnSync ETIMEDOUT, which is deliberately not treated as a checkpoint mismatch, so the move journal survived intact with all eight completed moves (score 17106 at move 8). Rerunning the same command replays those moves through the engine with per-move board and score verification and retries only the position that failed. Journal snapshot: runs/RUN-20260828T0700Z-d7-midgame-timeout/checkpoint-at-failure.jsonl.

Scripted-round results remain playground demonstrations, never tier evidence; no research seed range was touched.

A log entry is a narrative written by the contributors listed above. Run validity, scientific outcome and evidence tier live with the experiment and result records the entry refers to.