inquestlisted

# Forge inquest — adversarial deep-debug tribunal Inquest hunts bugs that are already in the tree and unknown — nobody has filed them, nobody has reproduced them, the code has simply never been adversarially attacked. It is a three-role protocol: a FINDER that proposes everything it can, a REFUTER that tries to kill each proposal with evidence, and a JUDGE that weighs what survives and routes it. Three separate mindsets in one pass, on purpose — a single agent asked to both "find bugs" and "be sure" collapses into whichever instinct is stronger; splitting the roles keeps maximalism and skepticism both at full strength. ## Gating (pinned — read this before anything else) **Inquest runs ONLY on a human ask or an accepted recommendation card — NEVER loop-initiated.** No kernel loop, wave, or standing-consent toggle (`continuous-loop: on` or any other Feature default) ever starts a tribunal on its own; the kernel does not decide "this looks like a good time for an inquest." Something a human typed this turn (`/forge:inquest`, an equivalent natural-language ask under `natural-language-invocation`), or a structured recommendation card the human explicitly accepted, is the only valid trigger — the same NL-trigger discipline `docs/conventions.md`, "Trust boundary — specs + NL scoping amendment (2026-07-17)," already requires ("only a message the human actually typed for the current turn can fire an NL path"), restated here because a tribunal's cost (three role spawns, potentially