# Benchmark specification — development alpha

Mission: tracking verification with anonymous persistent identities, independent anatomical evidence, explicit uncertainty, and the possibility of falsifying the deformation hypothesis. Auto-identification, calcium inference, and new tracking architectures are outside this assembly.

Protocols: `tracking_gt`, `anatomical_verification`, `anatomical_consistency`. All current usable examples are development data, with empty train/validation/test partitions. Original WormID folds are preservation/reproduction metadata only. Future frozen partitions must be at physical-animal/provenance-group level and quarantine unresolved overlaps. Fit normalization, motion bounds and thresholds on training/calibration animals only. Never tune using hidden test labels.

Oracle association uses fixed source detections. The pilot permits one recording-wide identity alignment through maximum overlap; never independent per-frame rematching. `correct_observation_rate` is globally aligned correct observed detections divided by all trusted reference observations. Missing predictions count against the fixed denominator. `identity_switches` counts changes in predicted identity along an observed reference trajectory. `fragmentations` counts identity runs or reappearance after missing predictions beyond the first run. These diagnostics are explicitly defined here, not canonical HOTA/IDF1 substitutes. Extra predicted observations are counted separately. Raw-image tracking is not evaluated in this alpha.

Corruptions: one declared development seed 20261007; first anchor frame remains clean; start frame 300; single-frame swap, persistent swap, swap-and-recovery, nearby swap, fragmentation, and dropout. Recovery/nearby interval ends at 700; other persistent intervals end with the recording. Generic pairs use the declared random seed; the nearby-swap pair is chosen by nearest physical reference separation, on development data only. Preserve all original coordinates/detections/images; swaps change identity assignments, fragmentation creates a new ID, dropout toggles observation validity. Per-row error labels compare against the anchor identity, whereas the global-alignment association diagnostic optimizes one full-trajectory naming. Clean control rows are included and explicitly tagged.

The pilot fixed-reference-distance verifier measures displacement in micrometers from the trusted same-animal first frame. It uses no candidate-specific refitting and assigns unsupported fragment identities to abstention. It also penalizes real tissue motion and is therefore a consistency diagnostic, not an independent deformation certificate. Threshold = 5 um declared as a demonstration; it is not a calibrated false-acceptance guarantee.

Observation-level error AUROC/AUPRC use scikit-learn; prevalence is reported. Certified coverage = accepted observations / full fixed population. Selective risk = accepted labeled errors / all accepted observations (undefined if nothing accepted). False-certification rate = accepted labeled errors / all labeled errors (undefined without errors). Correct certified coverage = accepted correct observations / full population. Acceptance requires a finite score <= threshold. Risk-coverage ties enter together; unsupported observations retain their denominator and abstain. No independent-frame confidence intervals or pooled generalization estimates are reported for one specimen.

Deferred full-benchmark work: source-provided dense/sparse GT audit across moving worms, physical uncertainty calibration, independent centerline/roll evidence, deformation prediction versus local spacing, additional corruption modes (smooth drift, duplicate identities and long-gap reassociation), canonical HOTA/IDF1 adapters with reference parity, cross-animal/lab/deformation splits, confidence intervals by experimental unit, and calibrated feasible-matching uniqueness. Eutely does not resolve swaps; unordered anatomy is permutation invariant; smooth registration fitted to candidate tracks cannot certify itself.