Scientific Method

Convene a model council: the same question answered independently across different models, dissent surfaced, cruxes routed to probes

Attack an asserted limit, ceiling, or claim with designed experiments

Run a full invention campaign: ideate past a limit, refute, build, measure against acceptance criteria, provenance-search, certify

Run a full scientific investigation: hypotheses, controlled experiments, verdicts, ledger

Create or update the persistent hypothesis ledger (EXPERIMENTS.md) for this project

Use this agent as one independent seat in a model council — the same question is posed to several seats, each running on a different model (pass a different model override per spawn), blind to each other's answers. The seat investigates with its own read-only probes and returns a structured position with evidence, calibrated confidence, and an explicit "what would change my mind". Spawn 3-5 seats in parallel for judgment calls: design decisions, interpretation of ambiguous results, risk assessments, go/no-go calls. Do not use a council to settle purely empirical questions — those go to experiments.

Use this agent to turn one asserted limit, claim, or candidate root cause into a rigorously designed experiment — without running it. It returns a falsifiable hypothesis with an explicit null, a written probe artifact, a pre-committed outcome→conclusion table, and what each outcome unlocks. Spawn one per hypothesis when fanning out a falsification campaign (the parent runs the probes serially for clean numbers). Use whenever an investigation has accumulated untested assertions ("the ceiling is X", "the daemon causes Y", "Z can't work") that need designed experiments rather than debate.

Use this agent as the area chair who closes a peer-review round. It receives the submission packet, all independent reviews, the author rebuttal, and any re-scores, then issues the final decision (accept / minor-revision / major-revision / reject) with camera-ready requirements. It weighs evidence quality rather than counting votes — one reviewer with a failed reproduction outweighs three approving skims. Spawn exactly one, after rebuttal, never before all reviews are in.

Use this agent as one independent reviewer in a peer-review round for an invention, design, finding, or paper-style writeup. Each reviewer gets the same submission packet plus ONE assigned lens (soundness, prior-art/provenance, reproducibility, significance, or fatal-flaw) and works blind to the other reviewers. Reviews are execution-grounded — the reviewer runs the Reproduce block, searches prior art, or re-derives the numbers depending on lens — and return structured scores plus an accept/revise/reject recommendation. Spawn 3-5 with different lenses after a finding survives the refute gate and before it is built, merged, published, or sent externally.

Use this agent to evaluate a finding, claimed root cause, performance claim, invention, or research result by attempting to refute it before judging what survives. It returns a verdict (confirmed / prototype / research / kill) with the refutation analysis, calibrated confidence, and a kill_reason when applicable. Spawn one fresh refuter per claim — the author of a finding must not referee it. Use proactively before acting on any finding, sending conclusions externally, merging a "fix", or relaying research-agent results the parent has not independently verified.

Use this skill whenever someone doubts a number, demands rigorous proof of a cause, or wants something invented and proven — any time a benchmark, metric, ceiling, or root-cause story must be verified rather than trusted, or a limit must be broken with a built, measured mechanism. Typical situations: a measurement looks suspicious ("2x faster but I don't trust it"); an incident needs its true cause before a decision ("the team blames X — confirm it, no guessing"); ANY production incident, outage, brownout, latency/error spike, regression, or "started failing" / "load suddenly went 5x" / "every component met spec but it still broke" report where the cause is not yet proven — especially when a recent deploy or change is suspected ("is the deploy to blame", "this coincided with the rollout"), since the obvious suspect is exactly what needs a control case; an asserted ceiling or "impossible" limit needs breaking or proving real; numbers need validation before publishing; a problem needs an invented solution that demonstrably meets acceptance criteria. Runs as a falsification-first campaign — hypotheses, predictions before measurement, controlled experiments, adversarial refutation, calibrated confidence, persistent ledger — with invention as the default continuation: surviving limits get attacked with designed mechanisms, built, measured against tuned baselines, and provenance-searched (claims demonstrated or searched, never asserted). Triggers: "use the scientific method", "prove it", "validate these claims", "verify, don't trust", "are you sure", "root cause this", "root cause and fix", "why did X happen / why is X slow", "diagnose this incident/outage/ brownout", "retry storm", "cascading failure", "what caused the regression", "I need real numbers", "falsify", "is it real", "invent", "break the ceiling" — and any incident or any challenge to an asserted number or cause, even unnamed and even when phrased as a routine ask. Use proactively before asserting any ceiling or root cause of your own. Do NOT use for ordinary feature work, refactoring, or one-file questions.

1 hook across 1 event