A local decision-model toolkit with candidate readouts, shared-prefix inference and development-set temperature fitting for Choice, Boolean and Score outputs; no output-token decoding is required.
Encodes state and question-specific candidates, then scores them through decision heads. Reuse mode requires explicit backbone support; calibrated temperatures are stored in checkpoints.
Provides training, calibration and evaluation tooling for local decision experiments. A new application still needs held-out labels and a coverage/error assessment.
Pinned README and model implementation reviewed. The site did not train or run a checkpoint; reported calibration and CPU results do not establish reliability on arbitrary production inputs.
Train and calibrate local decision readouts while keeping evaluation separate from demonstration data.
Material analysed：
model.py checks supports_reuse and rejects unsupported reuse mode. The performance benefit cannot be assumed for every backbone merely because the public API exposes mode='reuse'.
Inspect the supporting source ↗The README explicitly distinguishes distribution loss from temperature fitting on a development split. Published synthetic/demo numbers need matching data and protocol before comparison.
Inspect the supporting source ↗Use it to investigate local training and measurable uncertainty. Prefer task-specific held-out labels when deploying on a new domain; do not inherit a demo's calibration claim.
Inspect model.py's mode checks and compare train/dev/test commands in the README. Ensure calibration data is not reused as the final evaluation set.