meanrev-paper — performance dashboard

* Win%/PF greyed = <10 closed trades (noise). B trades rarely by design; B models cascade-entry fills stricter than A (12/14bp vs 5/6bp).

One engine, one operator, three $30k books running concurrently. This is the honest denominator for "did the research programme make money".

The alt book is where the tail lives: 54 disaster stops totalling −40,557, of which 34 are AFTER the 2026-07-23 freeze, worst −72.2%. Its peak mark equity was $36,861, so its drawdown from peak is 82.3% — not the ~17% a stale peak_equity field reported before 2026-08-12. A's majors sleeve has no stop at all, so its "zero stops" is definitional, not evidence of safety.

297 OOS trades — readable. Compare OOS rank vs the in-sample 99.8th pctile.

This card is the only CLEAN test — everything before the freeze is the window the 6 were picked on. Names + freeze time are hard-locked in code.

n=450 — DECISION POINT REACHED. Combined $-5,946 over 86d = $-70/day.

PRE-REGISTERED 2026-08-12: Step 0 found the book is 72% beta. The fade, net of the 5.0bp the bot actually charges, was +551 over 45 days — in-sample +1,127, out-of-sample −575. Every bootstrap CI includes zero and the sign does not hold between windows. Prediction: ~$+12/day if the fade is real, ~$0/day or negative if it was beta. AMENDED the same day, before any hedged data accrued: the first pass summed whole hourly bars and dropped the fill-to-hour and hour-to-exit stubs, which put the beta share at 67%, the cost at 11.6bp, and the IS/OOS signs the other way round. Redone per clip on the bot's own fill prices. A corrected input, not a criterion moved after seeing outcomes — the rule below is untouched. DECISION at n≥150: if the combined 95% bootstrap lower bound is below zero while A's is above it, the edge is beta and the strategy is dead standalone. This is an OVERLAY — it reads A's real positions and hedges them, so trade selection is identical to A by construction and the hedge is the only variable (unlike B, whose net cap also changed which trades were taken).

Worst% = single worst trade (max adverse excursion). Names with worst > ~5% and long holds are trending, not mean-reverting — the discriminator between the keep-6 and the drop-5.