What problem does it solve? Debugging sessions and ML experiments often drift into circular reasoning, repeated dead ends, and unfalsifiable guesses. This Skill structures the investigation as a bounded hypothesis loop with explicit confirmation and refutation criteria, a tried/rejected ledger, and stop conditions. ## Core Features & Use Cases - Falsifiable Hypothesis Generation: Produces candidate explanations with pre-declared confirming and refuting evidence, cost, and risk for each test. - Tried/Rejected Ledger: Records verdicts (confirmed, refuted, inconclusive, deferred) so ruled-out paths are never re-proposed. - Bounded Loop Control: Stops on confirmed root cause, metric hit, budget exhaustion, or diminishing returns, with escalation to a human when stuck. - Use Case: A model's validation accuracy suddenly drops. Use this Skill to list candidate causes (data leakage, distribution shift, config regression), define discriminating tests for each, run the cheapest high-information checks first, and converge on a confirmed root cause with evidence. ## Quick Start Use the hypothesis-testing skill to investigate why my model's validation accuracy dropped after the last data pipeline change and produce a hypothesis ledger with verdicts.
