The exit-learning loop, run at maximum effort until every survivor confessed

Seven months of farmer trades leaked profit at the exit, so we asked, twice and under locked rules, whether the exit could be learned. The first attempt trained on 686 real trades and captured 4.6% of the oracle gap out of sample against a pre-committed 17.3% break-even bar. The operator then ordered a maximum-effort search: six independent design families, walk-forward validation, a physically locked test window, every hypothesis logged to a ledger, and adversarial auditors instructed to rerun code rather than argue with it. Three candidates beat the random bar in validation. The audits then established that all three were the same defect in different clothes: each policy had access to when the trade's window actually ended, which is future information. Remove it and the best model collapses below random; make the simplest variant implementable and it scores worse than the exit logic it was meant to replace. Our own evaluation harness had also flattered everything, twice, including the random benchmark itself, and the write-up retracts the seductive headline it briefly produced.

Attempt one: reconstruction validated against the trade tape at r=0.999, gradient model on 80,545 bars, train capture 5.8%, test 4.6% vs the 17.3% bar; a fixed take-profit baseline scored 0.3% on the same test trades. Attempt two: 6 design lenses (microstructure at 200ms book resolution, volatility/funding regime, survival hazard, parametric ratchets, gradient-boosted meta-labeling, cross-asset), 239 validation evaluations on the ledger, top three by dollar-weighted validation capture 0.289 / 0.258 / 0.215 against a random-exit bar of 0.202. Nine adversarial audits: the meta-label model's dominant feature was fraction-of-window-elapsed (permutation importance 0.066 vs under 0.005 for every market feature); zeroing that single column and rerunning the full cross-validation collapsed it to 0.056-0.075, below the random median. The fractional-window rules consumed the realized window length directly; the implementable ablation (fraction of the training-median hold) scored −0.053, worse than the incumbent. The audits also corrected the harness itself: the matched null on the pooled validation trades is p50 0.178 / p95 0.237, and best-of-16 selection should face a max-of-16 null (median 0.238, p95 0.279), under which every headline in the search sits at or below the random median. Verdict under the pre-locked rule: champion set empty, test window never unlocked, question closed. The one implementable exit improvement ever measured on this tape remains a dumb fixed take-profit worth about +$48 over five months.