What problem does it solve? A Bot's charter and never-list are written once and rarely checked again, so the first evidence of a breach is usually a human noticing something was sent. This Skill replays a Bot's last N runs and produces evidence-backed answers about whether it stayed inside its approval boundary. ## Core Features & Use Cases - Never-list replay: Matches every numbered charter clause against quoted run artifacts, recording breach, clean, or unauditable per clause per run. - Four system checks: Verifies the routine fired (expected vs. observed runs with miss dates), the approval boundary held, memory shards carried between runs, and spend stayed under the cap. - Ranked breach report: Outputs a dated markdown eval with breaches ranked by blast radius, unauditable runs counted separately, and one evidence-supported charter change proposal. - Use Case: Before widening a Bot's permissions or connecting a new plugin, run a weekly audit to confirm its last 20 runs contained no unauthorized sends, writes, or publishes. ## Quick Start Audit the support Bot's last 20 runs against its current charter and list every never-list breach with quoted evidence.
