Cogent: CPU-First Typed Decision Models with Option Independence by Construction

Cogent is a family of four typed decision models (32M to 400M parameters) built on Ettin reranker encoders. Each takes a state, such as a support ticket, an invoice or an alert, plus a set of typed questions (choice, ordinal score or yes/no), and returns a probability distribution for each question. A request is encoded once under a tree attention graph, and a small reader head scores each option with no path from one option to another. In exact arithmetic this guarantees three properties: the state is encoded the same way whatever is asked, a question answers the same alone as in a batch, and choice and yes/no questions satisfy independence of irrelevant alternatives. A pre-registered ablation finds that this guarantee costs nothing measurable in Brier score. On a laptop CPU, five-question requests run at 1% to 18% of Laya-typed's p95 latency. On the typed-decisions benchmark, where all Cogent results are fitted (in-distribution), Cogent-Large beats Laya-typed on accuracy, Brier score and calibration after Holm correction. On JevBench, the smaller tiers trail Laya-typed and Large ties it. The paper reports its evaluation protocol, audits and negative results in full.