# Audit: seed + a1mc replays on 21 Sep-5 Oct tape (auditor, 10 Oct, #816)
Scripts read: /root/tm/seed30/{seed30.py,a1mc30.py}; sensitivity run on trades30.csv and a copy of a1mc30.py.
Verdict: replay leg PASSES as out-of-sample evidence; both tests stay verified=false (forward n 28 each, hand rebuild and sellability not done).

## Look-ahead
- seed: band filter uses the cap at the last print <= created+300 s, entry = first swap >= created+305 s, seed cap and 5-min volume only from swaps <= +300 s. No look-ahead found. Stops/pulls only count events after the entry block.
- a1mc: entry first swap >= first swap+300 s, cap read at that print, late buyers counted from swaps < +300 s. No look-ahead found.

## Robustness (my numbers)
- seed S replay n=132 mean 1.38x. Minus top 3 winners 1.30x, minus top 5 1.25x. First half of trades 1.44x, second half 1.31x. Not top-heavy; S beats B (1.00x) by 0.38x, matches claim. Days: 12 of 15 above 0.8x, 1.0x or better in 11.
- a1mc BASE n=279 mean 1.60x but 241/279 (86%) below 1x. Top 5 winners (34x,22x,21x,18x,17x) carry it: minus top5 1.22x, minus top10 0.95x (below break-even). LATEB minus top5 1.56x, so the sizing edge is real-looking but BASE alone is a lottery on ~10 coins. 136 of 279 exits are "time" marks at the last print, no sell impact on 17x+ marks.
- Seed tp is booked at exactly 5x*0.93 (no gap, no sell impact); optimistic for thin pools.

## Replay vs forward
- seed forward S 1.17x n=28 (replay 1.38x); S and L identical forward, so the volume filter was not binding yet. a1mc forward BASE 1.28x n=28 vs 1.60x, LATEB 1.38x vs 2.19x (n=28 each, LATEB gain far smaller forward). Direction agrees, magnitude lower: expect real edge below replay.
- "late buyers" uses swap recipients, may be a router; noisy (TM caveat). NOREPEAT not replayable.

## Open (why still unverified)
Sellability (v3 WETH pools: free eth_call sell sim not run here), outside sellers near exit, hand rebuild of 3 fills, forward n>=30 closed for the kill rule.