# Text and code relevance diagnostic

This frozen fixture contains 14 actual files and 48 English/German requests. It is a
small, agent-authored diagnostic, not a human-labelled production quality gate.
No inference has been performed against these labels.

`inventory.json` records exact source bytes, upstream immutable commits and
complete licence files. Twelve documents describe fictional organisations and
procedures; no customer information is used. Two complete Rust files come from
unrelated upstream projects. Local `.gitattributes` prevents Git newline
conversion of source and licence files.

Calibration uses bicycle warranty, bakery labels, apartment maintenance and
aho-corasick. Held-out uses museum loans, irrigation, ferry operations and
semver. Topic families and upstream repositories are disjoint. Each bilingual
intent pair stays within one split. Positive requests include contrasting
nearby procedures; absent requests require a specific missing artifact or
implementation. A related topic alone cannot fulfil such a request.

The independent agent reviewer should read only `blind-review-input.json` and
its referenced source files, then return candidate file IDs, an empty set for
an absent artifact, or an explicit ambiguity. Do not read `author-labels.json`
until that review is saved. No vector, search ranking or threshold is available
to either annotator. Disagreements must be resolved with written reasons before
a final immutable label lock is created.

The independent AI review is now saved and agreed with all expected file sets.
`fixture.json` combines the reviewed queries and labels; `fixture.lock.json`
records their hashes and the UTC freeze time before any inference. Review history
preserves encoding repairs and wording clarifications before that freeze.
**Do not evaluate held-out queries until a policy or threshold has been fixed
using calibration only.** Do not adjust that policy from held-out results.

No media is included: this corpus cannot calibrate music, images, video or
spoken-passage abstention. Existing media benchmarks remain separate. A held-out
split made by the same author is weaker than independent human annotation and
larger real customer libraries.

Run `python fixtures/relevance-v2/verify.py` for source integrity and split
structure. That check proves consistency, not retrieval quality.
Run `python fixtures/relevance-v2/test_verify.py` for isolated corruption,
duplicate, split-leak and lock regressions, including verifier execution with
Python optimization enabled.