searchi/fixtures/mixed/README.md

Mixed All-files diagnostic v1

This is a frozen, small diagnostic, not a production quality score. The first genuine CPU mixed run is recorded in MIXED-QUALITY.md; its failures remain visible and v1 labels remain unchanged.

The corpus contains 32 actual files: six searchi implementations, four searchi technical documents, two complete pinned llama.cpp source files, and twenty existing licensed media files. Source snapshots retain exact original bytes; fixture.json records upstream paths, commits, SHA-256, sizes, and licenses. The llama.cpp copies come from committed upstream objects, excluding local runtime patches. Its MIT notice is in LLAMA-LICENSE.txt. Media files are referenced from ../media/assets rather than duplicated. Their source URLs, licenses, independently observed labels and attribution remain in the media fixture.

There are 91 frozen queries: English/German implementation and document pairs, 65 reused media queries, four additional absent-content queries and four pinned upstream implementation queries. Code queries run in both All files and Code scope. Documents about the implementation coexist with the code, providing realistic competing content. Speech, music, images and videos compete with both. The case asking for code that chooses speech search is intentional: it can expose audio-intent routing that suppresses a relevant implementation.

Labels were written by one annotator before mixed inference. Existing media results had already been observed in their separate evaluation, so this is not an untouched held-out media benchmark. searchi sources are the application's own repository; the two llama.cpp files add external code, but do not establish quality across unrelated real projects. Never adjust v1 labels after seeing mixed results. Corrections require a new fixture version, explicit reasoning, and a new pre-inference lock. A second blinded human review remains open.

Run python tools/mixed-eval/evaluate.py --verify-only to verify labels and all source/media bytes without invoking a model. See the evaluator instructions for real runs.