This directory is the versioned fixture the decided eval grounding benchmark scores. It is a dev/CI surface, not a RAC artifact corpus — nothing here is part of the product knowledge under rac/.

• corpus/ — a small, schema-valid fixture corpus (decided validate exit 0), spanning all five artifact types, modelling a fictional collaborative editor ("Aurora"). It includes:
  • a supersession chain: decision-token-expiry-v1 (Status: Superseded) is superseded by decision-token-refresh-v2 (Status: Accepted). The superseded member is the canonical must_not_return hard negative — a query about the current policy must surface v2, never v1.
  • distractors: e.g. decision-rate-limiting ("token-bucket") shares the word "token" with the auth decisions but must not be confused with them.
  • an ambiguous pair: design-sidebar-navigation and design-mobile-navigation both concern navigation; a specific query must disambiguate to the right one.
• queries.json — the scored query set. Each case names exactly one tool (search_artifacts or get_related), a query (a search string, or the artifact id to look up for get_related), a category, the relevant ids, and an optional must_not_return set.
• eval-config.json — the gate's floors and tolerance. Gated metrics: overall.p_at_1, overall.r_at_5, negative_violations, and each query category's p_at_1 / r_at_5. Per-tool figures are diagnostic.
• baseline.json — the committed metrics baseline, written by decided eval --update-baseline (human-only; CI never rebaselines).

rac eval                  # human-readable scorecard
rac eval --json           # full scorecard JSON
rac eval --check          # CI gate: exit 0 pass / 1 regression / 2 usage error
rac eval --update-baseline  # human-only re-baseline

The initial floors come from the first green run, on which every category scored 1.0. Overall floors are the fixed targets (p_at_1 ≥ 0.90, r_at_5 ≥ 0.95); per-category floors sit a margin below their calibrated values, with the committed baseline minus the tolerance (0.02) providing the tight per-category regression guard. When the corpus or query set legitimately changes, re-run decided eval --update-baseline and commit the new baseline alongside the change.