# Search quality: source evidence and correction

Reviewed September 6, 2026 by Maguyva Team. [Machine-readable record](https://maguyva.ai/evidence/search-quality.json).

This is an audit of implemented classification rules and the source of a historical claim. It contains no retrieval benchmark results, customer queries, or statistical confidence estimate.

## What the fixture thresholds mean

The language engine's `AccuracyMetrics.status` classifies freshly calculated fixture accuracy as GREEN when precision is at least 0.95, recall at least 0.99, F1 at least 0.97, and the fixture has at least 20 expected edges. Additional curation rules cap the result at YELLOW when more than half the expected cases were automatically added, curated F1 is below 0.90, or the module-source share exceeds 0.50 without an exemption. Serialized manifests can preserve an explicit status override.

The 20-edge floor is a minimum for known test cases. It is not a calculated confidence interval and does not establish real-world accuracy. A language's enabled AST, symbol, or graph flag does not assert that its fixtures are GREEN. The previous methodology wording incorrectly implied that all enabled languages had passed a universal release gate.

Audited source revision: `6eb3d8c25437578b1365f127a8af52b3dde471a9`.

Source: `packages/maguyva/language_engine/src/code_language_engine/tools/audit/validation_manifest/metrics_models.py`.

SHA-256: `5be5629920550bc5993c664cf76dde93671aacfde9677fdb2df84f677f5d3683`.

The site's evidence test exercises the current classification boundary and curation caps against these published values. It checks the rule, not extraction performance.

## Correction: the historical zero-result reduction was a target

An earlier version of the multi-modal fusion article said the zero-result rate fell from roughly 15% to under 5%. We withdraw that achieved-improvement claim.

The December 26, 2025 redesign proposal lists roughly 15% as an estimated baseline and under 5% as a target. Its testing plan includes creating a benchmark query set. The implementation commit likewise describes the lower rate as a target. The subsequently auto-extracted decision, `DEC-SR-027`, recast that target as an achieved reduction. That decision record is not supporting measurement evidence.

No query set, sample size, before-and-after results, or measurement receipt supporting that reduction is supplied. Consequently, this record reports no achieved rate and no reproducible empirical improvement.

Historical source commit: `0cad65765b1f5d386c43e6943e2b8e7d9b163081`.

Proposal path at that commit: `packages/maguyva/docs/refactoring/INTELLIGENT_SEARCH_REDESIGN.md`, including its quantitative metrics tables and testing plan.

Proposal SHA-256: `5d04e40a4f3b10666691cd49213a1ff8b971691009b33f00afc933290867888f`.

## Reproduce the source checks

These commands require a checkout with repository history; source paths and hashes identify the material audited here and do not grant repository access.

```sh
cd packages/maguyva/public_site
npx vitest run tests/scripts/search-quality-evidence.test.ts
```

From the repository root:

```sh
git show 0cad65765b1f5d386c43e6943e2b8e7d9b163081:packages/maguyva/docs/refactoring/INTELLIGENT_SEARCH_REDESIGN.md
git show --format=fuller --no-patch 0cad65765b1f5d386c43e6943e2b8e7d9b163081
```

Configured capability counts are documented separately on [Compatibility](https://maguyva.ai/docs/compatibility). They are configuration counts, not quality benchmark results.