TAP gate redesigned 2026-10-06 from a 30-day backtest (2,848 windows, 11,288 evaluations at 5/4/3/2 min, real Kalshi settlements): model confidence ≥80% hit 96.9% (n=3,918) — m=5: 94.5%, m=4: 96.2%, m=3: 98.4%, m=2: 99.9%.

Removed: the 85¢ price lockout (excluded a 97.4%-hit pool) and the 3¢ edge bar (filtered backwards: 72.6% with edge vs 81.0% without).

⚠️ High-confidence entries are often 85¢+ — risking ~85¢ to make ~15¢. The hit rate only pays with exit discipline (take-profit/stop), not lazy holds.

STRONG: |dist| ≥ 20 bps + conf ≥ 98% → ~90% hit, ~14% flip. SCALP: 2–5 min + conf ≥ 65% → 88.5%. Flipped positions won only ~38–41% — a confirmed flip is an exit.

Model is systematically underconfident late in windows. Coinbase spot ≠ official CF Benchmarks RTI settlement basis. One regime sample — the live paper ledger is the forward test.