technology
Research that stays in papers changes the conversation. Research that ships changes the outcome. Symbolics Lab builds working systems from its formal results — tools that make theoretical properties available to practitioners who may never read the underlying work.
Classification
Tercet: Typed uncertainty
Standard classifiers compress every prediction into a confidence score — a single number between zero and one. The compression is lossy. What is lost is the structure of the evidence: whether it converges on one answer, narrows to a few, or fails to discriminate at all. These are qualitatively different epistemic situations, and they require qualitatively different responses. A confidence score hides the difference.
Typed uncertainty replaces the score with a structural judgment. Every prediction returns one of three types: the evidence singles out one class (commit), the evidence narrows the field without deciding (review), or the evidence is insufficient (escalate). The type is not a threshold on a score. It is a property of the evidence distribution itself, derived from the lab’s formal work on distinction, paraconsistent logic, and policy-governed classification.
In benchmark evaluation across 18 datasets, the approach does not achieve higher overall accuracy. It achieves a different kind of result: when the classifier commits, committed accuracy exceeds 95%. Cases that would otherwise produce confident errors are flagged before they reach a decision-maker. The instrument reports its own limits.
Tercet
Classification API
Production API implementing typed uncertainty for any classification task. Scikit-learn compatible. Every prediction returns CERTAIN, PARTIAL, or UNCERTAIN.