What problem does it solve? Changing an evaluation golden Q&A corpus from weak or single-source evidence silently corrupts every future eval round, A/B comparison, and re-judge. This Skill enforces a verification workflow so golden answers, key facts, avoid clauses, and grader notes are only changed after cross-class corroboration, with disputed or unverifiable facts encoded honestly instead of pinned. ## Core Features & Use Cases - Truth-domain classification: Classify each claim as real-world, corpus-grounded, or freshness-sensitive to pick the correct verification standard before editing. - Multi-source corroboration matrix: Require agreement across independent source classes (official docs, source code, live APIs, web research, docs index, empirical execution) with per-claim verdicts recorded in truth.corroboration[]. - Verdict-based encoding and lifecycle: Encode confirmed, disputed, contradicted, or unverifiable facts differently, manage case lifecycle states (proposed, active, quarantined, retired), and run lint, compile, and register checks before closing a change. - Use Case: An eval case claims a Stellar RPC numeric limit that a live probe contradicts. Use this Skill to triangulate the limit across docs and source code, update the golden with an asOf date and evidence trail, sweep sibling cases for consistency, and pass the CI gospel-change lint. ## Quick Start Use the golden-truth skill to verify and update the disputed numeric limit in eval case eval/qa/corpus/battery/<category>/<id>.json with corroborated evidence.
