# h3n native replay v2 (testmgr2, 10 Oct) - answer to auditor2 FIX #667
Script /root/tm/replay/h3n_nat2/h3n.py (changes vs h3n_nat): (1) `reserve` = net ETH flow of swaps strictly BEFORE the entry swap (old value included the entry swap itself, which inflated the 0.5+ buckets); (2) reserve impact per side, factor (1-0.04/(R+0.04))^2, R floored 0.01; (3) exit clock = real entry +595 s (fill delay now a cost), not first_swap+600. Costs 0.3% fee + 3% slip as before.
Reserve timestamp answer: old = cumulative up to and including the entry swap (not later data, but includes the entry swap); fix above.
## Arm A, n233
| bucket (pre-entry reserve ETH) | n | mean | median | mean cap10 | zeros |
|--|--|--|--|--|--|
| all | 233 | 3.66 | 3.80 | 3.23 | 56 |
| <0.1 | 63 | 0.22 | 0.00 | 0.22 | 46 |
| 0.1-0.5 | 149 | 4.93 | 4.95 | 4.35 | 8 |
| >=0.5 | 19 | 5.50 | 4.49 | 4.77 | 1 |
| >=0.1 (candidate rule) | 168 | 4.99 | 4.94 | 4.39 | 9 |
| fill<=10 s | 81 | 4.51 | 4.96 | 4.42 | 2 |
| fill>10 s | 152 | 3.20 | 1.29 | 2.59 | 54 |
- Old 9.7x medians were mostly the entry-swap inclusion. Honest figure: ~5x mean/median on pools with >=0.1 ETH before entry; <0.1 ETH (no prior liquidity seen) loses and must be skipped, a rule knowable at entry.
- Fill delay: slow fills (>10 s from init) are where zeros sit; on-time fills 4.5x with 2 zeros. Forward paper must agree before verified; sellability still waits on Ben's $5 test.