These entries are not GitHub issue bodies and do not block strict MVP. Promote one only after scope review, then assign the appropriate optional priority.
SHOULD#
- Temperature scaling using validation data only; compare against unscaled calibration.
- Confusion-pair commentary grounded in measured evaluation results.
STRETCH#
- One compact chart generated only from real measured metrics.
POST-WEEKEND#
- Docker packaging. Local
uvand Make remain the strict-MVP interface. - Multilingual evaluation that does not alter the English BANKING77 training contract.
- Per-class thresholds selected on validation evidence.
- Stronger out-of-distribution evaluation with a declared external dataset.
- Quantization or ONNX export.
- Batch prediction endpoint.
- Ticket-system adapter.
- Drift analysis.
- Human-review workflow.
- Cloud deployment.
- Load testing.
Initial treatment: keep all optional work in this single parking-lot document. Creating draft Project items or issues now would add noise and imply commitment.