Deepan Wadhwa writes about OpenDecision, a Python package that makes structured semantic decisions using a zero-shot NLI model (~400M ModernBERT-large) instead of a generative LLM. It exposes three typed primitives—Choice (select from a set), Noul (binary predicate), and Score (ordered rubric)—and serves them locally via FastAPI with a TypeSafe SDK-compatible endpoint.
- Inspired by TypeSafe's Jev "System One Model" announcement; the author previously built similar fraud-detection logic for a healthcare client.
- Choice uses two complementary NLI "compilers" and falls back to a third adjudication pass when they disagree, all on the same model.
- On TypeSafe-adapted benchmarks: 84.3% Choice accuracy, 85.0% Noul, 0.375 MAE on Score; on its own 125-case holdout: 86.4%.
- The returned "confidence" value is a concentration measure over the probability distribution, not a calibrated correctness estimate.
- Inspired by TypeSafe's Jev "System One Model" announcement; the author previously built similar fraud-detection logic for a healthcare client.
- Choice uses two complementary NLI "compilers" and falls back to a third adjudication pass when they disagree, all on the same model.
- On TypeSafe-adapted benchmarks: 84.3% Choice accuracy, 85.0% Noul, 0.375 MAE on Score; on its own 125-case holdout: 86.4%.
- The returned "confidence" value is a concentration measure over the probability distribution, not a calibrated correctness estimate.
