What problem does it solve? Research teams often misread experiment outcomes: they retrofit explanations to match results, draw conclusions from weak evidence, or keep tweaking without converging on a decision. This Skill enforces disciplined evidence interpretation so every run either shrinks the set of live explanations or produces a clear kill/pivot/continue verdict. ## Core Features & Use Cases - Baseline-Expected-Actual Comparison: Recovers pre-registered baselines, expected ranges, and mechanisms before looking at results, then tabulates deviations per metric. - Causal Chain Diagnosis: Locates which link broke—validity, intervention, mechanism, target, or value—using a decision tree in references/causal-chain-diagnosis.md, with a filled failure-diagnosis example. - Statistical Verdicts & Convergence Accounting: Selects analysis methods by experimental design, reports effect sizes with uncertainty intervals, and records how each run updated hypotheses, live explanations, and the next decision. - Use Case: After a training run misses its expected +6pp gain, use this Skill to determine whether the intervention signal never appeared, the mechanism failed to transfer, or the metric decoupled from the research goal—then get a concrete discriminating next test. ## Quick Start Analyze the results of experiment E05 against its pre-registered expectations and tell me whether to continue, pivot, or stop this direction.
