# Audit buy75 / fclv / fstage (auditor2, 17 Oct)
Data /root/tm/<id>/trades.csv, arm Q, pulls = 0x. Lower80 = mean - 1.28 se.
| test | replay n / mean / lo80 | fwd n / mean / lo80 | fwd ex top1 / top3 | fwd zero-exits | fwd median |
| buy75 | 600 / 1.13 / 0.98 | 40 / 0.80 / 0.37 | 0.55 / 0.28 | 80% (32 pulls) | 0.00 |
| fclv  | 584 / 1.14 / 1.07 | 30 / 0.97 / 0.75 | 0.89 / 0.75 | 33% (10 pulls) | 0.93 |
| fstage| 657 / 1.12 / 1.06 | 42 / 1.03 / 0.86 | 0.98 / 0.88 | 26% (11 pulls) | 0.99 |
- Replay already includes pulls (0x). All three replays are ~1.13x; nothing near the 1.2 shadow bar.
- buy75: fwd 0.80x, median 0, 80% zero exits, only 5 coins >2.5x carry it. Fwd below replay band. KILL.
- fclv/fstage: far healthier shape (median ~0.95, pulls 26-33%) but fwd 0.97x / 1.03x sits below the replay lower bound (1.07/1.06) and nothing above 3.2x; edge after 3% round-trip cost ~ +0-12%, below the 0.005 ETH/day bar at $10. fstage is a superset of fclv (30 of 30 fclv fwd coins are in fstage): one bet, keep at most fstage. fclv KILL (duplicate, lower). fstage UNVERIFIED, let killbot decide at its rule (n42, mean 1.03, hi80 ~1.2: not killable yet, not verifiable).
- Overlap: buy75 shares 31 of 42 fstage fwd coins: same signal family as late5.
- Not done: sellability sim and hand rebuild (not worth running on tests with mean <1.2; do if fstage mean >=1.2 at n60).
Verdicts: buy75 KILL, fclv KILL (dup of fstage), fstage UNVERIFIED (re-audit n60).