# Parity fixtures (reduced-v1) The 54 recorded production search responses, reduced to what the parity harness (`tests/search_parity.rs`) actually grades: the recorded status, the full sorted `(dataset, name)` identifier set, and the first-ten-in- recorded-order top 10. Error shapes carry their diagnostic body instead. - Recorded from `https://genenetwork.org/api3/search/` on 2026-10-05 (`per_page=50000`, `page=1`; see the recording script's parameters); 14 fixtures re-recorded (Q02, Q14, Q19, Q20, Q22, Q23, Q26, Q32, Q33, Q34, Q40, Q41, Q44, Q45) on 2026-10-07, see below. - The raw payload corpus (~121 MB) deliberately stays OUT of git; these reduced files (7.2 MB) are what grading consumes, and the harness's loader accepts both formats (top-level array = full fixture, object with `"format": "reduced-v1"` = reduced) in one directory. - Regenerate from a full corpus with: `python3 reduce_fixtures.py . ../../tests/search_parity.rs` — the shape table is parsed from the harness itself so the two cannot drift; the generator fails loudly on any hit missing its `(dataset, name)` identity. The 2026-10-07 re-recording closed the truncation gap and made the zero-hit shapes discriminating where possible: - Q02 (`insulin`, gene) was re-recorded through a verified pagination walk: 50000 + 7632 = 57632 hits, pages proven disjoint — no fixture in this corpus is truncated any more. - Q26 (`chr:2`) was proven to match >1.1M documents and was re-shaped to the bounded `chr:m` (9374 hits), which exercises the same `chr:`→`XC` boolean-filter grammar. - 12 shapes that previously recorded zero hits (Q14, Q19, Q20, Q22, Q23, Q32, Q33, Q34, Q40, Q41, Q44, Q45) were re-shaped to real values, so clause (b) of the parity metric bites on 36/36 shapes instead of 24/36. - Q04, Q11, Q12, Q21 honestly remain empty-vs-empty: the deployed index carries no positional data (phrase/NEAR/ADJ can never match) and the `description:` (XD) field is unpopulated. They exercise parser acceptance only until the indexer is fixed.