NOVA

Language models answer exact questions with total confidence and perfect grammar — and get them wrong: on our 250-question reserved set, Muse Spark alone scored 0.540 on stories with negations, refunds, duplicate amounts, and ambiguous scope. NOVA solves that — without changing the model. NOVA is a proof-first reasoning API that makes a language model trustworthy on exact questions — the kind with negations, refunds, duplicate amounts, and ambiguous scope, where models confidently produce wrong answers in perfect grammar. How it works: a question goes through three cooperating parts that share one concept representation. (1) The concept engine parses the story into event frames — who gave what to whom, in what order, with negations and polarity preserved — and runs a deterministic ledger over them. (2) The symbolic reasoner derives the answer with a step-by-step proof trace, or abstains and states why the question cannot be determined (ambiguous actor, a refund that matches no transfer, a missing fact). (3) The companion handoff sends the model the original question verbatim — never a rewrite — alongside what NOVA verified: the arithmetic and proof, the relevant word senses, what was and was not understood, and an honest unknown where one exists. The model then writes the final answer grounded in that material, and can still overrule NOVA when the question text disagrees (that happened in testing, and the model was right). In the builder's own words, NOVA is 'more like the language model companion that helps the language model — like more information than what the language model could supply itself.' It never replaces or redacts the question, never forces the model to defer to a verdict, and never claims verification it does not have. No fine-tuning, no weight changes — the capability is added by architecture, alongside an unchanged model. Measured results: on a reserved set of 250 exact stories, Muse Spark alone scored 0.540, rising to 0.880 with the NOVA companion attached (fuller pipeline 0.988) — same model, unchanged weights. Latest live head-to-head on 40 fresh questions: 39/40. Deterministic core: 250/250 on frozen and fresh sets, 391 unit tests passing. Live API: https://nova-proof-api-118169638073.us-central1.run.app — POST /reason takes a plain question story and returns the answer, the events, and the proof. NOVA is honest about its limits: it holds no general knowledge. On a 60-question textbook science test it verified 0 of 60 on its own because the model already knows that material; its value concentrates where models are weak — exact, multi-step, trap-laden reasoning.