An ensemble that explains itself

A classifier that’s right but mute is a hard sell. Reviewers don’t just want the label — they want to know why. The trick is to pair a fast model that decides with a talkative one that explains, and to keep those two jobs strictly separate.

Cheap, local, and good enough

For narrow text classification, TF-IDF plus embeddings into a logistic-regression head is fast, runs entirely on your own hardware, and — trained on tens of thousands of labeled examples — is genuinely accurate. No per-prediction API call, no data leaving the box, no latency surprises.

Let the LLM talk, not decide

The tempting, risky move is to let an LLM do the classification. The safe move: the ensemble decides, and the LLM writes a justification grounded in the input — quoting the exact phrases that drove the label. Because the explanation is generated after the decision, it can never quietly change the verdict.

Measure before you believe

Every modeling choice — how much data, which embedding head, ensemble versus a single model — earned its place by moving a number on a held-out set. “Feels better” is not a result. Recall at a fixed precision is.

Write down the dead ends

The experiments that didn’t work are worth documenting. The next person — usually future you — saves weeks by learning that the fancy finetune underperformed the boring linear head.

Let the cheap model decide and the expensive one explain. Trust lives in the seam between them.