Back to Model
S

SafeDecide

Model Oct 18, 2026

A 140M typed-decision model and calibration library that return probabilities with an automate or escalate flag.

Visit primary link 2 sources attached
SafeDecide's author-published demo escalates an approve answer at confidence 0.52 below a 0.70 threshold; the benchmark answer is reject.
Author-published demo for a held-out invoice example. The model's proposed answer differs from the benchmark, and its gate escalates the decision. This single example is not an AITrending test or evidence of universal error prevention. Full frame preserved. Source: Yotam Braun / SafeDecide
self-hostedverification

What happened lately

SafeDecide timeline

SafeDecide scores typed yes/no, choice and ordinal questions with a small cross-encoder. Its library calibrates a threshold on labeled data and returns automate or escalate, rather than treating every confident answer as actionable. The released 140M checkpoint is based on mmBERT-small and labeled Apache-2.0.

Python 3.11 or newer is required for the documented library installation. PyTorch, ONNX and browser-oriented exports are described. The shipped statistical guarantee assumes data exchangeable with the calibration set; int8 exports need fresh calibration because their logits change. Evaluation uses teacher-derived labels and does not establish suitability for consequential automated decisions about people.