What problem does it solve? Agents acting autonomously on compliance tickets need a consistent way to decide when they may approve or deny changes on their own versus when they must defer to a human reviewer. This Skill defines the confidence framework that makes that decision deterministic and auditable. ## Core Features & Use Cases - Gate thresholds: Maps confidence scores to allowed actions — autonomous approval/denial at 0.90+, approval plus notification at 0.75–0.89, advisory comment only at 0.50–0.74, and human-review flagging below 0.50. - LLM-context-rule cap: Caps confidence at 0.70 (advisory-only mode) when no LLM context rule pod exists for the target entity or its parent. - Ambiguity rule: Defaults to advisory comment-only behavior when guidance is ambiguous or signals conflict, with transparency about what a human should review. - Use Case: A policy update relaxes coverage.overall.minimum from 0.8 to 0.5; the agent diffs the policies, finds an explicit context rule forbidding the relaxation, computes 0.92 confidence, and autonomously submits a denial with a quoted justification. ## Quick Start Ask the agent to compute a decision confidence score for the pending ticket and state which gate action it permits.
