Brief
Ship BoW vs frozen-mini-transformer/API sentiment baselines with accuracy/F1 + two error slices (negation, domain slang).
Dataset
Sentiment split; review-id leakage trap graded via ablation.
Starter guardrails (88)
- Slots: BoW run, transformer/API run, metric table, slice reports, id-ablation.
- API runs must log cost/latency per call batch.
Deterministic checks
Macro-F1 >= 0.78 on hidden slice; id-ablation present; slice reports cover the two required phenomena.
100-pt rubric
- BoW tuning honesty: 20
- Transformer/API comparison fairness: 20
- F1 bar met: 15
- Error-slice depth (2 required): 25
- Id-leak ablation: 15
- Cost/latency logging (API path): 5