Refutation-by-majority (M13)

High-stakes findings (audit findings at or above a severity threshold, structural claims destined to be acted on) are not accepted from a single agent. Independent skeptical agents are dispatched with the explicit mandate to REFUTE the finding against the actual code; a finding survives only if it withstands a strict majority.

Why

Single-reviewer verdicts inherit the reviewer's blind spots and a "reviewing to approve" bias. Refutation-framing plus independence is the cheapest known de-biaser; majority survival bounds both false positives (R4 — the "hexagonal architecture is fiction" incident was the fiction) and rubber stamps (R2).

Dispatch protocol

1. Threshold. Default severity threshold is high (critical + high). Lower only for audit modes that explicitly demand it.

2. N is declared, not improvised. Read N from .claude/agent-dispatch-matrix.json refutation_skeptics for the active tier:

   tier	N
   XS	1
   S	1
   Standard	3

   Solo XS/S = 1 skeptic (the red-team IS the skeptic). Standard = 3. Gated-review raises the floor to the full vertical set (matrix tier_verticals.Standard).

3. Independence. Spawn N read-only skeptics (registry ladder per .claude/AGENT_REGISTRY.md; rule-50 read-only clause makes them legal in parallel without worktrees). Each skeptic receives:
   • only the finding text + its citations — no sibling verdicts, no orchestrator opinion;
   • the mandate: "your job is to REFUTE this claim against the actual code; return UPHELD only if you fail to refute it."
4. Persist. Each skeptic's return is an E1 agent-return envelope (role:"skeptic" + refutations[]), piped through node scripts/record-agent-return.mjs --task '#NNN'.

5. Marker. The orchestrator writes .arbiter/evidence/agent-returns/<task>/refutation-required.json at dispatch time — before the skeptic verdicts come back, as a numbered step of the review phase. The gate below adjudicates only when the marker exists, so leaving it to post-hoc discretion turns the gate into one that never fires. When the review produced no findings at or above the threshold, still write the marker with an empty findings array: that records the threshold was evaluated, instead of leaving "no marker" ambiguous between "nothing to refute" and "step skipped".

   { "task": "#NNN", "threshold": "high", "skeptics": 3, "findings": ["f1", "f2"] }

   findings = the ids requiring refutation (the acted-on set). The skill refuses to conclude while scripts/check-refutation-verdicts.mjs is non-zero.

Majority rule

UPHELD > N/2 ⇒ the finding survives and may be acted on. Otherwise it is demoted to info and logged — never silently dropped (M14 discipline: a refuted finding is still a recorded finding, just not an actionable one).

The loop: hop until nothing above low survives (INV-145, CANON-24)

One pass is not the protocol. After the skeptics report and their findings are fixed, dispatch a fresh round against the fixed tree — the new hop attacks the fixes, not the original claims — and repeat until no finding above low remains unaddressed. Empirically (#2480) the second hop destroyed the first hop's fixes twice: a gate that passed round 1 was still scoring a fully commented-out document 12/12.

Each hop is independent by construction: new agents, disjoint scopes, prompts naming the ARTIFACT and never the previous round's reasoning. Reusing a skeptic, or handing one the last round's rationale, converts refutation into confirmation.

Prefer the cross-model seat. When crossModelReview is configured (arbiter.json, provider e.g. codex), the hop runs there: a different model is a materially stronger skeptic than the same model in a different costume, because the blind spots are not correlated. The same-model fan-out below is a fallback, and it is weaker in exactly that way.

A hop that cannot run is recorded, not skipped. Model unavailable, rate limit, cross-model seat offline: self-probe the round's questions and write "degraded": true on the marker. The gate accepts it and says DEGRADED on every run. Never file a self-probe as an independent round.

Gate

scripts/check-refutation-verdicts.mjs (advisory runWarnCheck at L2+, promoted to runCheck at gated-review) adjudicates two axes over the same envelopes:

1. Majority — every finding in findings must have ≥ N skeptic verdicts and a strict UPHELD majority. Stops a phantom being acted on.
2. Severity floor (INV-145) — every finding the skeptics majority-UPHELD at critical/high/med must appear in findings. Stops the loop ending while something real is open. Severity is the highest any skeptic assigned. Below quorum or majority-REFUTED never blocks: one false alarm must not hold a wave hostage.

Marker present ⇒ both axes run, including when findings is empty — a round that addressed nothing while a high finding stood is precisely what the floor catches. No marker ⇒ the gate passes vacuously — which is exactly why step 5 makes writing the marker (empty findings included) the orchestrator's unconditional duty at dispatch: the gate can only adjudicate what was declared, and /ship and /drain both name this step in their review phases.

When a mechanical matcher — rather than this gate's structural check — has to decide whether LLM-authored text matches a ground-truth item, the sampled-audit-then-judge protocol in docs/internal/METHOD/ADJUDICATION.md applies.

Tier right-sizing

Fan-out never exceeds min(--max-parallel, nproc-2) (wave-drain cap). Solo never pays N=3 for a typo fix — N=1 is the red-team. Standard pays 3; gated-review pays the full vertical set. N is a declaration in the dispatch SSOT, not a runtime guess.