Retrieval encoders retooled for one-pass decisions: prior-fused heads hold up with scarce labels, cross heads win on full data

Biomedical retrieval sentence encoders can be turned into typed decision models that return, in one forward pass, a probability for each option so code can act, defer or escalate on a threshold. Existing System One models were general-domain and started from masked-language-model or decoder checkpoints, with no biomedical version and no test of whether retrieval training is a better starting point. Across six parent-retriever pairs, the cross head reaches 66–73 seen-task points with top-label ECE below 6 after temperature scaling (MedCPT-Query excepted), against 75.0 for fine-tuned Laya-large at roughly three times the size.

The conversion decides whether retrieval training helps: the prior-fused residual head adds a cross score on top of the retrieval prior, while the cross head encodes state and hypothesis jointly with a two-layer transformer. Retrieval similarity need not encode negation, ordinal judgement or calibrated truth, so whether retrieval training transfers to typed decisions was previously unanswered. Over 15 pair and data-size comparisons, retrieval training is significantly better for PFR in ten and worse in one, but better in one and worse in five for the cross head, which also beats PFR by 1.3–2.4 points on all five shared objectives.

The released RLCD recipe trails plain cross-entropy, and the deficit comes mainly from its reward normalisation. RLCD mixes cross-entropy with a policy gradient against noise-perturbed logits and had been the training recipe for open System One models. In the matched grid PG trails cross-entropy by 3.0 points (PFR) and 2.5 (cross head), with its estimator inflated 3.6–15-fold; PG-LOO with a leave-one-out baseline recovers 2.5 and 1.5 points, and after temperature scaling no objective is clearly better calibrated.

Next, the four conversions can be applied to a reader's own sentence encoder: start with a cross head and cross-entropy for per-request decisions, switch to PFR when low option-order sensitivity matters or the encoder's mask-slot states are degenerate, and use the bi-encoder B when a corpus must be screened against many questions. With scarce labels, prefer PFR, since retrieval training helps there most consistently. Long clinical notes can be tried with retrieve-then-decide, though most strategies stay near chance for base encoders.

MEDLINE-S1 labels come from NLM indexing and are weak, so absence need not mean false; a few training inputs recur in calibration and test splits, and the authors report a 0.3–0.4 point drop when those are removed but did not retrain or refit temperatures. Low ECE on seen tasks does not imply calibration on new tasks, cross-head option-order sensitivity grows under shift, and thresholds should be validated on in-domain samples. The clinical tests are small, come from one hospital system and use surrogate names, long-note strategies remain close to chance for base encoders, and the released model is trained on public data only.