What problem does it solve? During data analysis, debugging and iteration can silently drift into unauthorized changes — dropping outliers, switching estimators, or redefining the estimand — producing results the user never agreed to. This Skill draws a hard line between decisions an analyst may make autonomously and decisions that must stop for explicit user approval. ## Core Features & Use Cases - Decision classification: Distinguishes checkpoint-required changes (design, estimand, sample, specification, metric, any number the user has seen) from autonomous fixes (mechanical bug repairs that restore the agreed computation). - Structured checkpoint protocol: Name the decision, show the evidence, present at least two options with tradeoffs, give a recommendation, and wait — never implement past the gate. - Locked-document gate mechanics: Shared sign-off rules for planning artifacts like the framing brief, pre-analysis plan, causal Design Card, structural model card, and Prediction Spec. - Use Case: While debugging a difference-in-differences estimate, you discover a recording artifact that a triple-difference would absorb. Instead of silently upgrading the design, the Skill forces you to present options and wait for the user's choice before re-estimating. ## Quick Start Ask the assistant to apply analysis checkpoints while executing your analysis plan so it stops and asks before changing the design, sample, specification, or any reported number.
