Why an AI-discovered strategy can still be rejected immediately.

An AI-discovered rule can look excellent on paper and still be unusable in real time. If future information leaked into the inputs, or the data contract cannot be reproduced at the actual decision timestamp, the result is not a promising strategy. It is a failed candidate.

The first question is not, “How much did it make?”

The first thing to check is whether each feature was actually observable before the decision point. Performance can look deceptively strong if timestamps are coarse, post-processed labels leak into features, or later-corrected data is treated as though it had been available in the past.

AI can discover rules quickly, but it does not remove the responsibility to verify them.

The larger the search space, the more combinations will fit the past by chance. That is why candidate generation and validation are kept separate. We do not say that “AI found a profitable strategy” until it survives out-of-sample testing, walk-forward validation, realistic costs, and point-in-time constraints.

A good-looking result still needs explicit rejection criteria.

If we find look-ahead leakage, unrealistic execution, excessive cost sensitivity, unreproducible source lineage, or rules that were changed after the fact, the candidate is rejected regardless of headline performance. The failure mode is kept as a constraint for the next experiment.