epistemic

Use when a research idea, vague comparison, or surprising observation must be turned into a falsifiable, prereg-ready hypothesis before any benchmark, eval, training, or judge-backed work begins.

Use when starting a research, eval, benchmark, or empirical comparison session - explains how to activate Epistemic skills, choose stage skills, and enforce research gates before any experiment work.

Use when starting or continuing empirical/ML research, evals, benchmarks, or any "is X better than Y?" claim — the umbrella mechanism that enforces hypothesis → preregistration → baseline → experiment → statistical rigor → falsification → kill-or-ship → verification, with gates you self-enforce.

Use when a claim, result, or draft compares your system against an external baseline, published paper number, or competitor result and you need a sourced local reproduction before quoting it.

Use when a preregistered hypothesis is ready to run and you must generate provisional evidence under the locked method without contaminating headline outputs.

Use when a hypothesis has reached a real decision point and you must choose KILL, PIVOT, RECOMMIT, REFINE, or SHIP from repository evidence.

Use when a hypothesis is concrete enough to freeze the experiment contract, lock the judge and compute environment, and scaffold the run before any experiment-shaped command executes.

Use when completed experiment outputs exist in `experiments/{id}/smokes/` and the agent must decide whether an inferential claim is statistically defensible before falsification review or promotion.

Use when about to state, circulate, merge, or publish any research result, comparison claim, or confirmed conclusion.

Use when a numerical claim, comparison claim, or headline result is about to leave `smokes/` or be written into `RESULTS.md`.

Quarantine and explain any result that diverges from the most recent prior result on the same benchmark by more than 15% before it leaves `smokes/`.