Typed Decision Models: An Early Evidence Audit and Evaluation Checklist

Typed decision models return a probability distribution over options supplied by the caller, and software tends to act on that number as if it were a guarantee. We measure what it licenses, comparing the open-weight Laya, in five checkpoints, with Jev 1.13, available only through an API. A probability licenses a statement about groups of decisions, not about one decision, and a gate that promises an accuracy among the cases it decides alone keeps that promise only with a lower bound checked on data like its traffic. Across 190 exchangeable evaluations, a threshold chosen by point estimate breaks the promise in 147, Learn then Test keeps it in all of them, and a Beta lower-bound rule decides more items on average than Learn then Test in 128 and fewer in none, without its guarantee. Calibrated in English, Jev keeps a promise of 95% in 8 of 51 locales. Ordinary typing noise breaks a promise calibrated on clean text, and, in exploratory runs, training with noise raises the noisy accuracy with no loss on clean text. In four further domains, a fine-tuned Laya needs from 250 to 8,000 labels to be non-inferior to Jev, and does not get there on LIAR2. Under drift, the preregistered claim for a Bayesian retraining loop is not supported: the loop avoids needless retraining where the change does not degrade the model, at a cost equivalent to recalibration alone, and re-qualifying the threshold in every window keeps the promise more often where it does. On the monthly series of M3, neither decider predicts the direction of the next value better than a constant answer. Confidence separates right from wrong answers with an AUROC from 0.80 to 0.96, but it does not select faithful explanations, and exact coalition games of single decisions show what an explanation can still say through an API that rounds to two decimals. We release the frozen answers with a datasheet, a promise check and the code.

A typed decision model answers a fixed question about an input by returning a probability for each of several caller-defined options. Each option carries a short label and a written definition, which is where a developer states the rule the model should apply. Jev introduced this interface for routing, moderation and triage, open implementations followed, and the same operation occurs whenever a language model is used as a classifier by scoring label strings. We study the open implementations, whose weights we can inspect and patch, and ask whether the probability follows the definitions or the labels. A preference for the label we call option-label bias. Across four open-weight typed decision models, three ways of reading an answer from a Qwen2.5 backbone, eleven classification tasks and PolicyBench, a synthetic routing suite we introduce in which the rule appears only in the definitions, the answer is mostly the labels. Deleting every definition leaves accuracy unchanged (laya-td: 0.8559 against 0.8487), although those definitions support 0.7971 on their own, and renaming the options to A and B raises accuracy by +0.1511 [+0.1377, +0.1646]. One system, von, is unaffected, and the two code bases differ in one expression: laya writes each option as "{label}: {definition}", while von writes only the definition. Changing that expression in both directions, with no weight changed, makes all three laya checkpoints exactly invariant (+0.0000 [+0.0000, +0.0000]) and creates the effect in von, whose accuracy falls from 0.8511 to 0.2281 when a label contradicts its definition. Earlier work attributed this failure to the constrained decision head these models use in place of a text decoder; our results locate it in the prompt rendering. We give a two-call test that tells a practitioner which case applies to their model, and measure what four mitigations are worth.

Typed decision models (TDMs) return a probability over caller-defined options, and Jev is described by its vendor as calibrated. But in deployment class frequencies change, so calibration measured on a fixed test set need not hold where thresholds are applied. Existing evaluations leave two gaps: the gains of vendor decision training have not been separated from those of the backbone under an identical readout, and TDMs have been tested only at the test set's class distribution. We ask whether decision training improves a model over its own backbone, and whether a TDM's calibration and thresholds withstand prevalence (label) shift as well as a fine-tuned classifier's. We compare three TDMs (Jev, Tev1-4B, Nimble-9B), their base models and two trained classifiers on identical decisions from five public datasets under simulated shift, with 35 paired tests fixed before any test was computed. First, decision training is not uniformly beneficial: Nimble-9B, which differs from Qwen3.5-9B only by a LoRA adapter, is better calibrated on intent tasks but 12.0 Macro-F1 points worse on financial sentiment. Second, under moderate shift Jev's expected calibration error (ECE) and threshold risk grow significantly more than a fine-tuned DeBERTa's only on financial sentiment, by 1.7 and 2.5 percentage points; averaged over 2,000 shift scenarios (exploratory), the gap is at most one point on every dataset. Third, EM prior correction with the model's mean development prediction as its prior needs no labels and lowers Jev's ECE under shift by 2.7 and 1.8 points on the two datasets where it is highest, although labelled recalibration lowers it far more. The three TDMs we tested are neither uniformly better than their backbones nor uniquely fragile under shift; for calibration, their main weakness is the miscalibration they already have without shift. We release all inputs, code and outputs.