Type-Safe Is Not Error-Free: A Constrained Decision Head Follows the Option Name, Not the Rubric Bound to It
Swapping option names from 0/1 to no/yes reverses Jev decisions, dropping AUC from 0.94 to 0.23.
The paper finds that schema-constrained decision models can return valid options while following the semantic polarity of option names rather than the rubric bound to each name. On 1,200 workflow decisions, renaming the two options from 0/1 to no/yes changed 70.4 more answers per hundred and shifted AUC from 0.94 to 0.23. Across four predicates the effect was at least 7.4 times larger than a neutral-name control, while the type-error rate remained 0%. Random character-string names restored the neutral regime, and a mean-pooling readout flipped 4.1 times less often.
- Renaming 0/1 options to no/yes changes 70.4 more answers per hundred.
- AUC falls from 0.94 to 0.23, reversing the decision ranking.
- The effect is at least 7.4 times the neutral-name control.
- Type-error rate stays 0% even when decision accuracy collapses.
Full article301 words · extracted from arxiv.org · click to collapse
Typed decision models are built for settings where model outputs are consumed directly by software. Instead of generating free-form text, they return a decision over a predefined set of options. By construction, every output conforms to the required schema. Yet this guarantee does not tell us whether the model interprets the options as intended. We study Jev and two Jev-like models with open weights by changing how option names are assigned to rubrics. Each option consists of an option name and a textual rubric that defines what the option means. We change only which option name is assigned to each rubric; the question, state, rubric wording, and set of option names remain exactly the same. On 1200 workflow decisions with task-specific rubrics, renaming the two options from 0/1 to no/yes changes 70.4 more answers per hundred (95% CI: [67.6, 73.1]) and shifts AUC from .94 to .23, revealing a systematic reversal in the decision ranking rather than simple uncertainty. The same operation has little effect with neutral option names. This pattern holds across all 4 predicates, where the effect is at least 7.4x larger than under the neutral control, and becomes stronger as the number of options increases. The effect also depends on the read-out geometry: a second model family that mean-pools over the full option span flips 4.1x less often. The hosted model exhibits the same behavior: the swap changes AUC from .8146 to .5806 and produces 24x as many answer flips as its test-retest floor. In contrast, replacing the option names with random character strings returns all model families to the neutral-control regime without reducing accuracy. The failure therefore depends on the semantic polarity of the option names rather than on the renaming operation itself. Across all conditions, the type-error rate remains 0%, even when decision accuracy degrades substantially.
Text extracted automatically; images, tables and formatting may be missing. Original: https://arxiv.org/abs/2609.26758