Live decomposition of the 2026-07-23 drawdown (equity 62.6 -> 31.1 in one
day, 43 fills, avg hold <1h despite a 4h min hold): both exit guards
compared thresholds meant as price moves against leverage-multiplied
margin PnL%, so raising leverage 5x -> 10x silently halved every trigger
distance:
- throttle bypass: -5% 'stop' unlocked AI closes at a -0.5% price wiggle
(ETH -0.5%/44min, AAVE -0.9%/45min, MU -2.4%/14min all sailed through)
- drawdown monitor: 'profit > 5%' armed at +0.5% price, then market-closed
winners on a 40% giveback (SNDK +0.6%/35min, NVDA +0.3%/37min) — every
winner strangled at breakeven while losers realized fast
Fix: evaluate both guards on price-basis PnL (margin PnL% / leverage).
Throttle thresholds keep their documented price meaning at any leverage;
the drawdown monitor now arms only after a real +5% price move. Peak-PnL
cache stays margin-basis for prompt display consistency.
Replay simulator updated to the same price-basis semantics.
Replay recorded AI decisions from decision_records under alternative
risk-control and throttle parameters, then search the 15-dim space with
Optuna multivariate TPE.
- extract.py: parse decision cycles, intents (incl. throttled ones), 15m
OHLCV and per-cycle prices out of input prompts into CSVs
- simulate.py: replay engine mirroring auto_trader_throttle.go semantics
(margin-based PnL%% thresholds, intra-candle SL/TP trigger orders,
opens/hour + reentry + margin gates); live-config replay reproduces the
real account curve (-83%% sim vs -85%% actual)
- search.py: TPE search (train/holdout split) + baselines + importances
Findings: train-window optima do not survive holdout (overfit); only
2/800 trials are positive across all 3 time folds. Dominant lever is
min_confidence (importance 0.79) — the edge problem is decision quality,
not risk parameters.
