Investigate

Reduce uncertainty enough to make the next decision well. The goal is not information collection; it is a justified belief update.

Work

1. State the uncertainty. Write the question and how its answer could change what happens next. If it cannot affect a consequential decision, reconsider the investigation.
2. Establish the evidence already available. Avoid repeating prior work.
3. Maintain competing hypotheses. For causal questions, name plausible alternatives before committing to one. Include measurement or environment error when credible.
4. Seek discriminating evidence. Prefer cheap, high-information observations that make hypotheses diverge: reproduction, traces, logs, profiling, history, controlled experiments, minimal prototypes, primary documentation, targeted tests. When a performance number will drive a decision, follow references/performance-evidence.md to prove the measured work happened and the comparison is valid.
5. Update explicitly. Reject contradicted hypotheses, strengthen supported ones, and introduce new hypotheses when evidence requires it.
6. Reproduce when useful. For defects, obtain a reliable reproduction when practical. For design uncertainty, prototype the smallest risky assumption rather than debating it abstractly.
7. Stop at decision sufficiency. Do not seek certainty for its own sake.

Return

• Question
• Existing evidence
• Hypotheses considered
• Discriminating evidence gathered
• Finding and confidence
• Remaining uncertainty
• Decision implications

Failure signals

Watch for confirmation bias, one-hypothesis debugging, evidence merely compatible with a favorite explanation, research that cannot change action, and fixing before understanding when the root cause matters.

Done when

The consequential uncertainty has been reduced enough that remaining uncertainty is unlikely to alter the next decision, or further investigation is demonstrably not worth its cost.