Agents¶

Five specialized agents fan out under campaigns. Each has a narrow contract and a structured return format the orchestrator consumes raw. You can also invoke them directly by name.

Agents table¶

Agent	Color	Tools	Role
`experiment-designer`	blue	Read, Grep, Glob, Write, WebFetch	Design one hypothesis + probe + outcome table — does not run it
`refuter`	orange	Read-only (no Write/Edit)	Try to kill a finding first; return confirmed / prototype / research / kill
`council-member`	purple	Read-only (no Write/Edit)	One independent council seat: position, evidence, would-change-my-mind
`peer-reviewer`	yellow	Read-only (no Write/Edit)	One blind, lensed, execution-grounded review with scores
`meta-reviewer`	red	Read, Grep, Glob	Area chair: evidence over votes; final accept/revise/reject

experiment-designer¶

Turns one asserted limit, claim, or candidate root cause into a rigorously designed experiment — without running it.

When to spawn: one per hypothesis when fanning out a falsification campaign. The parent runs probes serially so measurements do not contaminate each other (device contention, cache state, cross-talk).

Returns, in order:

HYPOTHESIS — falsifiable claim, prediction, explicit null H0
EXPERIMENT — probe file path, how to run it, expected runtime, controls applied
OUTCOME TABLE — outcome → what it proves → next action (all plausible results)
UNLOCKS — what becomes possible if the assertion falls; what search closes if confirmed

Design discipline: name the suspected flaw in how the original assertion was produced (single-stream artifact, hot cache, derived formula, vendor default). Prefer the cheapest probe that fully discriminates the outcomes.

Used by: /scientific-method:investigate, /scientific-method:falsify.

refuter¶

Evaluates a finding by attempting to refute it first, then judging what survives. Fresh context is the point — the author of a finding must not referee it.

When to spawn: one fresh refuter per claim, before acting on any finding, sending conclusions externally, merging a "fix", or relaying research-agent results the parent has not independently verified.

Verdict enum:

Verdict	Meaning
`confirmed`	Survived refutation; evidence chain complete; safe to act on
`prototype`	Promising but needs exactly one more measurement — name it
`research`	Real but speculative or not feasible here yet
`kill`	Refuted, physically impossible, unbuildable, or no real gain (`kill_reason` required)

Classic holes it hunts:

Per-component win presented as end-to-end win
Correlation without a control case
Derived/spec-sheet number presented as a measurement
Missing baseline (regression could hide inside the "improvement")
Internal contradictions between sources
Claims above a measured physical ceiling

Used by: /scientific-method:verdict, invent pipeline, end of investigate campaigns.

council-member¶

One independent seat on a deliberative council. Several models get the same question; seats are blind to each other. The value is an independent path to an answer — dissent is the point.

When to spawn: 3–5 seats in parallel for judgment calls (design decisions, interpretation of ambiguous results, risk assessments, go/no-go). Do not use a council to settle purely empirical questions — those go to experiments.

Ground rules:

Probe before opining (read-only access to code/config/data).
Separate fact from judgment; name the probe that would settle unmeasured facts.
Commit to a position; "it depends" without naming what it depends on is a non-answer.

Returns exactly:

POSITION: <one-sentence answer/recommendation>
KEY EVIDENCE: <2-4 observations/probes with refs>
REASONING: <causal chain from evidence to position>
STRONGEST COUNTERARGUMENT: <best case against your own position>
CONFIDENCE: <0-1; 0.90+ only if grounded in direct evidence you checked>
WOULD CHANGE MY MIND: <specific observation that would flip you>
UNRESOLVED EMPIRICAL QUESTIONS: <facts + exact probes>

Used by: /scientific-method:council. Orchestrator aggregates by evidence not votes; factual cruxes become ledger hypotheses with probes.

peer-reviewer¶

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 and works blind to the others.

Lenses:

Lens	What it executes
`soundness`	Logic chain, controls, confounds, causal ladder
`prior-art/provenance`	Actual search log; closest found vs claimed novelty
`reproducibility`	Run the Reproduce block; check provenance header
`significance`	Does the gain matter end-to-end?
`fatal-flaw`	Always assigned — hunts the single blocker

Reviews are execution-grounded: run the Reproduce block, search prior art, or re-derive the numbers depending on lens. Return structured scores plus accept/revise/reject.

Used by: /scientific-method:peer-review (3–5 in parallel after refute gate, before build/merge/publish).

meta-reviewer¶

The area chair who closes a peer-review round. Receives the submission packet, all independent reviews, the author rebuttal, and any re-scores. Issues the final decision.

How it weighs:

Evidence beats votes. One reviewer who ran the Reproduce block and watched it fail outweighs three approving skims.
Rebuttal against reviews, not fluency. Blocking weaknesses answered only with new evidence or applied revisions.
Disagreement is information. Factual splits become named probes in the decision's requirements — not adjudicated by the chair.

Decision enum: accept / minor-revision / major-revision / reject, with camera-ready requirements. Maps straight to the ledger (accept → CONFIRMED; reject → falsification log with DO-NOT-RE-ATTACK).

Spawn exactly one, after rebuttal, never before all reviews are in.

Used by: /scientific-method:peer-review.

Skill¶

The scientific-method skill is the method itself. It carries:

The full campaign loop (hypothesis → ledger)
Artifact templates (references/artifacts.md)
Multi-agent campaign patterns (references/campaigns.md) — design/execute split, invent pipeline, council and peer-review contracts
Anticipation generators, rigor checks, causal structure, open-records playbook

It auto-triggers on phrases like "use the scientific method", "prove it", "validate these claims", "root cause this", "no guessing", "falsify", "invent", or any challenge to an asserted number, limit, or cause — and proactively before the model asserts a ceiling or root cause of its own.