# a1: mc5 exit audit re-run + serial relaunch #642 (task #2836, 17 Oct)
Script exr2.py = exr.py with holdout moved to 15-16 Oct (capture.py, 101 closed live mc5 trades x v3live swap path, 4h hold, 19 rugs priced as 0.0). Train = before 15 Oct (n=90), holdout 15-16 Oct (n=11, 3 rugs, live mean 0.31).

| rule (rug cost on) | train n=90 mean / w/o top3 | holdout n=11 mean / w/o top3 | all n=101 |
|---|---|---|---|
| base stop 0.5 | 1.73 / 0.91 | 0.61 / 0.31 | 1.60 / 0.87 |
| stop 0.3 | 1.83 / 1.01 | 1.48 / 0.19 | 1.79 / 1.07 |
| no stop, 4h | 2.64 / 1.76 | 1.35 / 0.02 (median 0.01) | 2.50 / 1.71 |
| delay 5 min then stop 0.3 | 2.09 / 1.29 | 1.48 / 0.19 | 2.03 / 1.31 |
| delay 30 min then stop 0.3 | 2.42 / 1.62 | 1.39 / 0.05 | 2.31 / 1.60 |

**Verdict:** direction holds on train and all (any later or wider stop beats stop 0.5 by 0.2-0.9x, rug cost ~0.1), but the holdout is only 11 trades and every rule is carried by 1-2 winners (w/o top3 0.0-0.2). 14 Oct pick (no stop 5 min, then 0.3) is still the safest improvement: holdout 1.48 vs 0.61 for base. Longer delays/no stop win on the mean but bleed to median 0.01-0.02. Not proven: needs shadow/dry run via exec; do not call it verified (holdout n<20). Caveats unchanged: swap-level, no slippage, stop fills worse live.

**#642 serial relaunch:** not replayed. Earlier serial/creator lenses (#305 P1M 0/81, #636, #598, #448) were killed; FOLIO's 5 relaunches are v4 native pools fmmc5 does not watch, and the prior-hit-by-symbol replay needs a 30-day v4 pool+symbol table (names.db covers symbols, no as-of prior-peak join yet). Ledger #642 stays idea; no data request added (v4tape ext and signer backfill are already open).