Skip to content

Commands

All commands are namespaced under scientific-method:. Run them from Claude Code after installing the plugin.

Commands table

Command Argument What it does
investigate <problem> Full campaign: hypotheses, controlled experiments, verdicts, ledger
falsify <claim> or all Attack an asserted limit, ceiling, or claim with designed probes
invent <problem> Invention campaign past a confirmed limit
verdict [claims] Adversarial REFUTE-first review before trusting findings
ledger [sync] Create or update the persistent hypothesis ledger
council <question> Multi-model council; dissent + factual cruxes → probes
peer-review <work> Blind lensed reviewers, rebuttal, area-chair decision

investigate

/scientific-method:investigate the API returns 500 under load but not in tests

Full falsification-first investigation. Contract:

  1. Cheapest probe first.
  2. Hypothesis ledger — every candidate as a numbered falsifiable claim with null; check DO-NOT-RE-ATTACK first.
  3. Predict, then probe — outcome tables before measuring; controls for causal claims; baselines for performance claims.
  4. Verdicts with evidence and calibrated confidence.
  5. Adversarial pass via the refuter.
  6. Loop until clean (dry pass, all verdicts filled).

If no problem is given, investigates the most recently discussed open question.

falsify

/scientific-method:falsify the bidirectional PCIe aggregate is 32.9 GB/s
/scientific-method:falsify all

Attack an asserted limit, ceiling, or claim. With all (or empty), sweeps docs/benchmarks/conversation for every "ceiling", "wall", "maximum", "can't", "impossible", and derived formula.

Derived numbers (X/2, spec-sheet values, library metadata) are prime targets — past campaigns falsified four of five asserted "physical" walls.

A confirmed limit is as valuable as a broken one: it ends a search honestly.

invent

/scientific-method:invent break the measured encode ceiling past 12.2% with acceptance ≥ +2pp over tuned baseline

Full invention campaign:

  1. Frame falsifiable acceptance criteria; measure the baseline first.
  2. Ideate with quotas and forced diversity.
  3. Refute before building; kills go to the falsification log.
  4. Build and measure survivors against pre-registered criteria.
  5. Provenance-search every surviving mechanism.
  6. Certify and ledger; high-stakes inventions go to peer-review.

verdict

/scientific-method:verdict
/scientific-method:verdict the outage was caused by the 14:00 deploy

The gate between "we found something" and "we act on it / report it externally". Spawns one fresh refuter per claim. Empty argument collects load-bearing claims from the conversation.

Report: verdict table, kills with reasons (appended to the falsification log), surviving claims with calibrated confidence.

ledger

/scientific-method:ledger
/scientific-method:ledger sync

Maintain EXPERIMENTS.md (or the project's incident/docs convention):

  • Create from the skill's artifact template if missing.
  • sync backfills untracked findings from the conversation.
  • Integrity check: demote verdicts without evidence; recalibrate 100% confidence with named unknowns; spot-check Reproduce commands.

Keep it append-only. The audit trail — including wrong turns — is what makes verdicts trustworthy.

council

/scientific-method:council should we ship the rewrite or keep patching the hot path?

Convene a model council for judgment calls, not empirical questions:

  1. Frame the question with evidence; do not include your own lean.
  2. Seat 3–5 council-member agents on different models, blind to each other.
  3. Optional out-of-family seat via COUNCIL_OOF_ENDPOINT (local OpenAI-compatible endpoint).
  4. Aggregate by evidence. Factual disagreements become ledger hypotheses with probes.

peer-review

/scientific-method:peer-review the encode-pipeline invention in EXPERIMENTS.md

True peer review after the refute gate, before build/merge/publish:

  1. Assemble submission packet (claim + all evidence + working Reproduce block).
  2. Round 1 — 3–5 blind peer-reviewer agents with distinct lenses.
  3. Rebuttal — answer with new evidence or applied revisions only.
  4. Re-score, then meta-reviewer area-chair decision.

Decision maps to the ledger: accept → CONFIRMED; reject → falsification log with DO-NOT-RE-ATTACK.