Kill Reports
Every killed strategy gets a detailed post-mortem. Which gate killed it, what thresholds it failed, the specific metrics at time of death, and the full audit trail.
What's in a Kill Report?
When a strategy is killed at any stage, the system records the complete context: the kill gate, the specific threshold that was breached, all metrics at the moment of death, the LLM critique (if at the interrogation stage), and every prior stage result. The result is a structured manifest.yaml file that serves as a complete forensic audit trail.
Example Kill Reports
Reason: Sharpe ratio 0.21 (threshold: ≥ 0.5). Only 12 trades over 5 years (threshold: ≥ 25). The strategy generated too few signals to be statistically meaningful. The hypothesis assumed overnight gaps would revert by market open, but in liquid ETFs the gap-fill rate is too inconsistent to build a systematic edge.
Verdict: KILL. The adversarial LLM (DeepSeek Reasoner) identified 3 critical issues: (1) look-ahead bias — the RSI was calculated using close price before market close; (2) overfitting — 7 tuned parameters for only 31 trades is a recipe for curve fitting; (3) regime dependence — the strategy only works in low-volatility bull markets and would have been destroyed in 2022.
Reason: In-sample Sharpe 1.89, out-of-sample Sharpe 0.34. Walk-forward decay: 82% (threshold: ≤ 50%). Classic overfitting — the spread relationship between SPY and IWM that existed in the training window (2020–2025) broke down during the test period (2025–2026). The strategy was fitting to a temporary correlation structure, not a persistent economic relationship.
Why We Publish Kill Reports
Every killed strategy teaches you something about what doesn't work in markets. A strategy killed at Stage 3 for insufficient Sharpe is a lesson in signal strength. A strategy killed at Stage 4 by an adversarial LLM is a lesson in bias detection. A strategy killed at Stage 6 by walk-forward validation is a lesson in overfitting. The graveyard is the curriculum — the survivors are the graduation.
Note: Builder and Institutional users get full access to kill reports including the generated Python backtest code, complete manifests, LLM critiques, and audit trails. Explorer users can view summary-level kill reports.