aif-grounded

Use a reliability gate that requires every factual claim to be backed by evidence from repository files, command outputs, or provided documents. If any claim lacks support, the assistant returns an insufficient-information result listing exactly what evidence is missing instead of guessing.

Classify the request as external-facts, then verify against authoritative docs, release notes, or logs before answering. If verification is impossible with available tools, return insufficient information and ask the user for the specific source excerpt or version output.

A confidence gate is a scoring mechanism where an answer is only delivered at 100 out of 100 confidence, meaning every claim is evidence-backed with no open unknowns. Below that threshold, the system outputs missing evidence and concrete steps to reach full confidence.

Refusal is appropriate for high-stakes domains like security, finance, legal, or medical questions, and whenever changeable facts cannot be verified. The Skill returns a structured checklist of missing evidence rather than a fabricated answer.

Yes, but patches are only applied once confidence reaches 100, meaning requirements are precise and build or test verification is possible. If the repo lacks a verification path, the Skill proposes the minimal validation needed instead of implementing blindly.