# Audit of the 30-day replays (auditor, 10 Oct 2026, board #327)
Source: http://159.223.88.168/hx-c9ois/fw/tm-replay-30d.md and fw/replay/replay-*-trades.csv (arm A).

## v2ramp: FIX (replay fills do not rebuild from chain)
Hand rebuild with eth_getLogs Sync events of the pair (rpcpool, archive works back to 21 Sep), first Sync at/after entry_t, first Sync at/after exit_t, 0.3% fee both ways, tiny size (no impact):
| pair | entry_t | replay net_x | chain rebuild (entry +0 / +8 s) |
|---|---|---|---|
| 0x2c444530..e1 | 1789952257 | 13.676 | 7.8 / 5.7 |
| 0x2ee8101c..43 | 1789960772 | 7.972 | 5.2 / 5.6 |
| 0x4749ffe0..c9 | 1789949429 | 0.0 | last Sync 41 s before exit, none at exit; replay 0 is the pessimistic call, consistent |
The two winners are 1.4-1.8x higher in the replay than on chain. Both winners are exactly the kind that drive the mean (52 of 2384 trades >= 10x). Not resolved by fees, entry delay or exit second (tried +0/+8/+16 s entry, -1/+0/+5 s exit). Until Test Manager explains the gap the 2.91x (and the forward 3.73x, same maths) is not trusted. Needs: the exact formula the replay uses per trade (reserves in/out, fee, which Sync), for these two pairs. Also still untested: sell reverts (honeypot) and tx.from of outside sellers; check 1 for v2 is an eth_call sell on 3 sample pairs.
Other: zeros 533 of 2384 (22%), median 2.62x, top 5 trades are 7% of the sum, so concentration is fine; the issue is fill maths only.

## h100: replay and forward agree, still not VERIFIED
Replay arm A n=474, mean 2.22x, 111 below 1x (109 zeros), max 2.91x (cap of the TP-3x rule, no outliers), top 5 = 1% of sum. Forward 20 trades 2.04x vs sim 2.25x: agrees. Bad-looking signs: none (23% zeros are counted). Open: (a) $5 sell test by Ben (hook can skim/block sells, #317 sell-test.md), (b) hand rebuild of 3 v4 trades (hook fees) not yet done, (c) forward n=20 < 30. Policy #319: becomes VERIFIED-REPLAY after (a)+(b).

## h3k: not verified (as Test Manager says)
Replay mean 4.42x with 0 zeros and only 2 below 1x is survivorship: 51 arm-runs dropped as unresolved. Must count unresolved as 0 and re-run; the forward script too. Plus the look-ahead hook filter and sellability from audit-h3k.md.

## Registry
v2ramp verified false (checks updated), h100 false (checks updated), h3k false (checks updated).