Noma: An Open System One Decision Model from a Sliced Decoder and Trained Heads

System One decision models answer typed questions about a state in a single forward pass, returning a probability for every option and no generated text. The class was introduced with Jev, a hosted model. We present Noma, an open-weight model in this class, and a measured account of its design: a pretrained 4B decoder sliced to its first 18 of 32 layers, trained decision heads that read one hidden state per option, a separately supervised abstain head, and a fast serving path. On single-pass decisions Noma scores 82.6% on a sealed, human-reviewed set that we also used to compare development runs (81.3% for a fresh re-training of the same design), with mean confidence about two points above accuracy. On two intent tasks absent from its training data it reaches 83.6% (150 options) and 70.1% (77 options) with no adaptation, and it answers in 15 ms at the median on one H100 with zero output tokens. Controlled ablations (20 runs at a 6,000-item recipe, paired per item) show which parts carry this. Slicing loses no measured accuracy (79.4% against 79.4% over three seeds each; 95% interval on the difference [-2.6, +2.7]). The abstain output is informative only when it is supervised (area under the ROC curve 0.82 against 0.46, one seed, 15 items). The option head can be small: one pointwise head about an eighth the size of the four-head listwise reference scores 80.2% against 79.4% (difference +0.8 [-1.3, +2.8]), and with all the training data the two designs remain within noise of each other. A single pass has limits: on 56 held-out items that need chained computation, every configuration with at least 18 layers scores between 44.6% and 62.5% (chance is 33.5%), the released model's confidence is 23 points above its accuracy, and 9.7% of multiple-choice answers change when the options are reversed. Code and weights are open.