FUTURE · SHORTLIST AND TARGETED TUNING

Select Strongest 1–3 + Targeted Tuning

Select only the strongest one to three models, then spend tuning budget on failure modes observed in the shared audit.

RECOMMENDATIONDIAGNOSTIC-DRIVENSHORTLIST ONLY

Resolution / small objects

Adjust resolution and the native small-object strategy only when size-slice evidence identifies the bottleneck.

Long-tail strategy

Use sampling, weighting, or class-aware augmentation only when rare-class evidence justifies it.

Crowding

Target density, matching, query, crop, or resolution behavior supported by crowded-scene diagnostics.

Localization

Tune geometry, regression, or post-processing only when localization error is material.

Schedule

Change duration or schedule from training-health evidence, not habit.

Counterfactual augmentation

Activate only when its specific context-shortcut or over-detection failure mode is observed.