Primitive of the day · 2026-10-02

rank-order-not-absolute-threshold

Cross-encoder scores are model-config-dependent raw logits (measured -11.4 to +7.8 on the default model) — consume rank order only, never threshold on absolute score.

When this applies

Downstream logic is about to consume cross-encoder relevance scores.

Preconditions

• Scores come from a cross-encoder predict/rank call
• Downstream logic can operate on relative ordering instead of a cutoff

Do this

Sort candidates by score and consume positions; if a cutoff is unavoidable, derive it empirically per model+corpus rather than assuming any score scale.

What you should see

Stable behavior across models whose score ranges and output activations differ.

How it fails if ignored

An absolute threshold silently dropping everything (or nothing) after a model swap; assuming sigmoid [0,1] output when the model's saved config sets Identity activation.

Do not use when

A model whose activation you have verified emits calibrated [0,1] scores AND whose calibration you have tested on your own data.

Kind: gotcha-fix. Part of the skill Build RAG reranker pipeline. Free to reuse in your own agent skills.