Keep sources attached to claims
When using AI to extract research questions, variables, or arguments, retain the original source and publication date. Check generated citations, data definitions, and factual claims against the source. An unresolved uncertainty is more useful than a confident explanation built on a missing reference.
Ask for the opposing case
After generating an explanation, ask what could invalidate it. Which market conditions break the signal? Does the idea survive higher costs? Could the data contain selection bias? This gives AI a useful role as a question expander without treating it as an authority over the evidence.
Translate language into explicit rules
Before backtesting, convert an idea into inputs, observation windows, exposure limits, and exit conditions. Review generated code for time alignment, missing values, cost handling, and exceptional cases. A program that runs successfully may still implement the wrong hypothesis.
Assign responsibility clearly
Models assist organization and explanation. People define the question, review the evidence, set risk boundaries, and make decisions. The public Aurenso lab uses transparent numerical rules. It does not call a language model or present demonstration outputs as a model’s actual market predictions.