# Calibration: harness vs live fills (auditor, 10 Oct 2026, task #372)
Source: fw/a3-live-audit.md (chain-only audit of wallet 0x0184ef9b, data to 8 Oct 16:56 BKK; last live fill) + fw/a1-live-vs-harness.md. Daily: this file is appended once a day, newest first.

## 10 Oct
Live fmmc5 fills, $10 size, 12 buys / 12 sells (6 Oct - 8 Oct):
- Model fill vs real fill, buys: fill/model 1.0000 on 12 of 12. Sells: 1.0000 on 11 of 12, one 0.856 (3 earlier swaps in same block, 15:35 7 Oct). Harness depth model is exact at $10; no systematic slip.
- Live x minus paper twin ($10), 12 closed sells: mean +0.008x, median +0.001x. Sum live 54.3x vs paper 54.2x.
  - Stops (10): live mean 0.411x vs paper 0.439x (-0.028x). Cause: gaps through the 0.5x stop (3 cascades: -0.26, -0.15, -0.02).
  - Time exits (2 winners): live 20.0x / 30.2x vs paper 19.4x / 30.4x (+0.58, -0.20). Noise, no bias.
- Entry delay (+5 min check to fill): 1-4 s on 11 of 12, one 8 s (WARN). Within the 3 s harness delay assumption +-.
- Gap is not fills. It is execution failures: 6 harness trades not taken live (a1-live-vs-harness: HTTP 429 cost a 19.5x winner, ~+$925 at $50; 2 GMGN no-hash; 1 confirmation required; 1 cap_out edge 245.5k vs 250k; 1 feed miss). Of ~21 attempted buys 3 FAILED.

## Verdict
Harness fill model for mc5 is CALIBRATED at live size ($10): error under 0.03x per stop trade, 0 bias on winners. Caveat: n=12, live size is $10, so depth effects of $50-$1000 are untested live. Hooked-v4 families (h3k/h100) have no live fills at all, so nothing to calibrate there; sellability still needs Ben's sell test.
Note: a3-live-audit last live trade is 8 Oct 16:56, while a1-live-match lists 14 closed to 8 Oct 23:00; not re-checked on chain today (no firstmin access by design). Next daily run: extend when a3 shows newer fills.