Closes a research campaign — audits integrity, synthesizes beliefs including negative results, and optionally prepares a human-reviewed promotion package. Use when a campaign stops, completes, or aborts; when the operator requests an integrity audit alone; or…
Places documentation in the correct home, trims duplicated rules, validates with verify_template.py, and archives or supersedes Agent Notes. Use when writing, moving, or auditing docs; when adding, auditing, archiving, or deleting Agent Notes; or when…
Chooses the cheapest sound verification method (tests, property-based tests, Z3, or Lean) and writes a proof plan with property, assumptions, model, commands, and unproved remainder. Use when a claim needs a test or a proof. Live Z3 or Lean binaries are not…
Shared research-campaign references. Use when writing or reviewing research skills, campaign files, experiment writeups, evaluators, proofs, or technical research prose. Read these files before inventing terms or SKILL.md sections.
Runs a campaign after setup — seals a baseline, executes a bounded trial loop, evaluates candidates, diagnoses crashes or invalid trials, updates research issues, and plans valley exploration when one greedy step cannot test a compound claim. Use when…
Forms a falsifiable hypothesis under lab/experiments/planned/ (Predictions before any run, no Results), runs the protocol, and writes up under successes/ or failures/. Use when proposing, running, or writing up an experiment, when a campaign needs a linked…
Classifies research work as spike, experiment, campaign, bug investigation, or design decision; creates a bounded lab/campaigns/<slug>/ tree; builds an invariant evaluator contract and lock input. Use when the operator asks to start research, scope unclear…