refutationlisted

# 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 v