Research question
What evidence should a strategy candidate survive before it is treated as more than an attractive backtest?
Principle
The useful adversarial question is not only “does it work?” but “what change in data visibility, baseline, cost, or implementation would make the result disappear?”
Method
The validation ladder keeps simple comparators alive: no-news and no-LLM variants, transaction checks, stale or shuffled-input diagnostics where appropriate, and turnover-aware evaluation.
Promotion boundary
A candidate is promoted only when data visibility, implementation correctness, baseline comparisons, and failure modes remain intelligible.
Limitations
No checklist can eliminate specification search, regime dependence, or undetected implementation errors. The method creates evidence and traceability; it does not certify truth.
Next question
Which test would be most likely to falsify the current interpretation?