# Audit buy75 + late5 (auditor2, 17 Oct)
Data: /root/tm/<id>/trades.csv, main arm Q, pull exits count as 0x.
| test | replay n / mean | fwd n / mean | fwd lower80 | fwd ex top1 | fwd ex top3 | fwd pull exits |
| buy75 | 600 / 1.13x | 30 / 1.07x | 0.51 | 0.74 | 0.38 | 22/30 |
| late5 | 462 / 0.96x | 21 / 1.75x | 0.90 | 1.31 | 0.72 | 14/21 |
- Median 0 in both (replay and fwd): money comes from 3-4 tail winners (5.5x, 5.7x, 7.4x, 10.6x). 19 of buy75's 30 fwd coins are also late5's coins, so the two are one bet, not two.
- late5 fwd 1.75x is outside its replay band (0.96x, 73% to 0); buy75 fwd 1.07x vs replay 1.13x is inside the band but lower 80% bound < 1.
- Not checked this run: sellability sim, hand rebuild (fills in these are pull=0x model). 
Verdict: both UNVERIFIED (verified stays false). buy75: no edge shown (1.07x, ex-top3 0.38x), expect KILL at n30 rule (fwd mean<1.0 after tail removal). late5: replay <1 so the fwd 1.75x is tail luck; re-audit at fwd n40, kill if mean <1.0 then. Neither is worth live size.