Target: report accuracy/F1 + error slices for a BoW baseline against a frozen-mini-transformer (or API) baseline on the sentiment split.
1. BoW baseline with teeth
TF-IDF + logistic regression, tuned C, honest validation. Weak-looking baselines usually mean weak tuning — beat it properly before claiming transformers helped.
2. Frozen transformer comparison
Frozen embeddings + linear head (or a cost-logged API baseline). Compare on accuracy AND macro-F1 (class imbalance) plus two error slices (negation, domain slang).
3. Trap: label leakage via id
Review-id ranges correlate with collection waves and labels. Drop/ablate id features; verify the delta before trusting any accuracy.
Recap
Tuned BoW anchor, frozen-transformer comparison on acc + macro-F1 + slices, and an id-ablation proving the numbers are real.