Trutina: Typed Decision Probabilities from a 4B Language Model with Letter Readout and Trainable Early Exits

Typed decision models answer a question about a state with a probability for every option instead of generating text. We present Trutina 4B, a decision model built on Qwen3.5-4B that reads the answer from the LM-head logits of the option letters at a fixed position, after a rank-16 LoRA fine-tune of all linear layers. Four design choices carry most of the result. First, presenting the question before and after the input raises zero-shot accuracy by 3 to 4 points on a backbone that is mostly linear attention. Second, a two-round tournament extends the letter readout to 255 options without new parameters. Third, affine exits at layers 12, 16, 20 and 24, distilled from the final layer with detached inputs, let a request stop at a confidence threshold and otherwise continue from where it stopped. Fourth, continued training with a high share of replay adds new capabilities without forgetting. We also report that a decision head newly trained on hidden states scored 74.0% on the JevBench public set, below the 79.7% of the untrained backbone's letter readout, which led us to abandon it. At full depth the model reaches 87.4% on the 231 JevBench public tasks and 74.65% accuracy (KL 0.156) on the zero-shot LocalLLaMA typed-decisions test set. The adaptive mode uses 21.6 of 32 layers on average at 86.6%, with a median latency of 14.0 ms on one RTX PRO 6000. Among six open 4B decision models run under one protocol, it has the highest accuracy on three of eight evaluations and is weaker on hard multi-step judgements. Weights and code are released under Apache-2.0.