# Focus local video search evaluation protocol v1

Version: 1.0.0 (2026-08-29)

This is an open data dictionary and set of blank tables for independently evaluating local video indexing and retrieval. It is not a Focus performance result.

## Required files

- `corpus.csv`: one row per rights-cleared source clip.
- `runs.csv`: one row per timed indexing run.
- `queries.csv`: pre-registered speech, visual, person, and scene queries.
- `qrels.csv`: human relevance judgments made before examining ranked results.
- `results.csv`: ranked results returned by each tested system.

Use stable IDs to join the files. Record failures instead of deleting them. Run at least three cold and three warm trials per clip. Publish individual observations as well as the median and range. Compute precision@5 and recall only when queries, qrels, and ranked results are all present.

## Reproducibility checklist

1. Use redistributable media and publish each source URL, license, and SHA-256.
2. Record app/build commit, operating system, device, chip, memory, power state, model versions, configuration, and cache state.
3. Keep the corpus and query set fixed across systems.
4. Separate indexing timing from retrieval-quality evaluation.
5. Publish raw observations, scoring code, exclusions, failures, limitations, and corrections.

The January 2026 Focus pilot cannot be reproduced bit-for-bit because its internal source clip is not redistributable and its original output omitted several environment fields. It should not be used as a product ranking or large-archive estimate.

License: CC BY 4.0. Corrections: contact@use-focus.com.