evaluation/protocol.md

# QT26-QC canonical evaluation protocol

## Submission

Submit one TSV row per `public_id` with columns `public_id`, `top1`, ..., `top20`.
Prediction values are FishBase 25.04 `canonical_taxon_key` strings. All 6,719
canonical IDs must occur exactly once.

If several submitted labels map to the same canonical key, duplicate canonical
keys are removed and later distinct keys shift forward. A top-k submission must
therefore supply at least k distinct canonical guesses after mapping. Invalid,
unmapped or blank predictions occupy the fixed denominator and score as wrong;
rejection and structural non-support are also errors.

## Metrics

- Primary: query-micro top-1 and species-macro top-1.
- Secondary: query-micro and species-macro top-5 and top-20.
- Paired comparisons: wrong-to-right, right-to-wrong and net-correct counts.
- Uncertainty: 20,000 target-species-cluster bootstrap replicates.

Query-micro weights every query equally. Species-macro first averages within
each target taxon and then weights taxa equally. Scores are fractions in the
JSON output; multiply by 100 for percent.

## Taxonomy mapping

Outputs from another label system must be mapped to FishBase 25.04 before
submission. Accepted names, synonyms and finer ranks that resolve to one species
should map to one canonical key. If no defensible mapping exists, retain an
unmapped token; do not drop the query.

## Reference results

`reference-baselines.tsv` records frozen COVER-Fish configurations as reference
baselines, not state of the art. The Fishial comparison in the paper is confined
to its shared 775-name task and is not a full-roster QT26-QC baseline.

## Example

```bash
python evaluation/score.py predictions.tsv --output metrics.json
python evaluation/score.py new.tsv --compare old.tsv --output paired.json
```