Deploying ML to an agent — what goes wrong

Your training data was BTC in a bull market. Today is a bear market. The model sees features it's never seen and produces garbage probabilities. Mitigation: retrain rolling-window. Watch for feature drift.

Your model needs bar-close features. In live trading you have the PREVIOUS bar's close, not the current one. Mitigation: use only features computable from bars that have fully closed.

Your test Sharpe ignored fees. Live trading pays 0.52% round-trip. If your average trade wins 0.3%, you lose money. Mitigation: include the fee in the training label — "is this bar's NEXT move > 0.52%?" not just "> 0%".

The PRD-STRATEGY-ONBOARDING.md pipeline is the right tool:

• Gate A backtest with realistic fee model
• Gate B testnet paper with LIVE data (catches latency bugs)
• Gate C mainnet probation with circuit breakers

An ML model is ONE strategy going through the same gates. It doesn't get special treatment.

For each of the three killers, name ONE mitigation (not the one from above). Multiple valid answers.