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:
- Cheapest probe first.
- Hypothesis ledger — every candidate as a numbered falsifiable claim with null; check DO-NOT-RE-ATTACK first.
- Predict, then probe — outcome tables before measuring; controls for causal claims; baselines for performance claims.
- Verdicts with evidence and calibrated confidence.
- Adversarial pass via the refuter.
- 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:
- Frame falsifiable acceptance criteria; measure the baseline first.
- Ideate with quotas and forced diversity.
- Refute before building; kills go to the falsification log.
- Build and measure survivors against pre-registered criteria.
- Provenance-search every surviving mechanism.
- 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.
syncbackfills 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:
- Frame the question with evidence; do not include your own lean.
- Seat 3–5 council-member agents on different models, blind to each other.
- Optional out-of-family seat via
COUNCIL_OOF_ENDPOINT(local OpenAI-compatible endpoint). - 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:
- Assemble submission packet (claim + all evidence + working Reproduce block).
- Round 1 — 3–5 blind peer-reviewer agents with distinct lenses.
- Rebuttal — answer with new evidence or applied revisions only.
- Re-score, then meta-reviewer area-chair decision.
Decision maps to the ledger: accept → CONFIRMED; reject → falsification log with DO-NOT-RE-ATTACK.