# Dev-floor fixtures (reduced-v1, dev-oracle) The 54 shapes of tests/search_parity.rs recorded from the SELF-BUILT xapian index (47,040 documents; public 2025-02-18 database dump + RDF caches; built with the production indexer — recipe: ~/spikes/index-self-build/README.md), reduced with the committed reduce_fixtures.py. These are DEV-TRUTH counts, not production counts: they grade the native engine and the pipeline against a reproducible, fully local index (the dev-oracle), independent of production access. - Recorded 2026-10-08 by tests/search_native_dev_corpus.rs (the recorder parses the harness table and records through the native engine; FATALs on mismatch). Regenerate with that test against XAPIAN_DB_PATH pointing at the rebuilt index. - The raw 128 MB corpus stays out of git, like the production one; these reduced files (5.2 MB) are what dev-floor grading consumes. - The production corpus (../parity-fixtures/) remains the grading truth for production parity; this corpus is the dev-oracle for engine mechanics and future re-baselines if the owner blesses the dev-oracle as the sign-off standard. - Data-model note (verified 2026-10-08): BXD traits appear under two dataset names (BXDPublish and GITrMetPublish — InbredSetId 1 has two PublishFreeze entries), so a trait Id can legitimately show up twice in one response with different dataset labels and identical LOD. These are distinct index documents, not duplicates: all 47,040 (dataset, name) keys are unique.