- commit
- 531c7c630902f87f2ac6091e18c584bf6e900477
- parent
- b53323b684bbb40b915d40c325544501ab3aafe4
- author
- dami <gazey55@gmail.com>
| M | PLAN.md | 30 | +++++++++++++++++------------- |
@@ -763,19 +763,23 @@ if slate is down. The app machine keeps its emergency field.
763763 The Codex agent's script commits every unknown cell of a revealed tile764764 as 0 in one batch, reads the rejections (which echo the value), commits765765 the complements, and never deduces anything: 723 tiles, 41k accepted766-cells, 20,557 wrong, top of the day's leaderboard. Within the stated767-rules — the tool text itself says a wrong binary value is full768-information — and exactly what a per-cell verdict invites. A Takuzu is769-self-verifying by its rules; the referee must not be a "check this770-cell" button.771-772-Decision to make (not shipped yet, it changes the live rules): commits773-are claims, verified per tile when the agent claims it complete; a774-failed claim resets the tile's claimed cells, costs reward (at least775-the tile's cells) and a cooldown, and reveals only "wrong", not where;776-wrong% on the leaderboard; the per-cell verdict survives only as a777-localhost practice knob. A genuine m-way ambiguity then costs at most778-m-1 failed claims, which is the price of a guess made real.766+cells, 20,557 wrong (33%), top of the day's leaderboard. Within the767+stated rules; the per-cell verdict invites it. Honest play measured the768+same day: 0.1-0.3% wrong.769+770+**Decided 2026-08-27: health.** Guessing bleeds you. Each agent has a771+health read from its wrong-rate over its last 2,000 committed cells772+(windowed, so an honest rate heals the damage on its own): under 1.5%773+healthy; 1.5-4% warning (flagged in observe and commit results, hue774+wilts in the viewer); 4% and above, after at least 300 committed cells,775+death — writes refused for the rest of the day, territory stays, group776+mates continue, the leaderboard shows "died at rev N", tomorrow is777+fresh. Rationale: an honest 3-way tile costs up to ~1.3% (two wrong778+cells per ~150), the exploit runs at 33%, a blind prober dies after779+~250 cells. Per-cell verdicts stay: they are what the rate is measured780+on. The prompt describes health as a fact of the field, not as a list781+of prohibitions (agents on the host cannot start games, so the prompt782+does not tell them not to).779783 780784 ### Status of the v2 items under this goal781785