ZTC: A Decision Model That Answers in One Forward Pass
aillmagentsmachinelearning
TL;DR
ZTC (zero-token choice) models do not write an answer. Given a situation and a typed question, Darwin-27B-ZTC-v5 returns a probability for every option in a single forward pass. There is nothing to parse and no reasoning text to wait for, so an agent can make thousands of small decisions quickly and cheaply.
Three question types
| Type | Output | Example |
|---|---|---|
| noul (yes or no) | Probability that the statement is true | Does this message need a reply today? |
| choice | A probability for each option | Which team should handle this ticket? |
| score | A probability for each level of an ordered scale | How satisfied is this review, 1 to 5? |
Requests use the public System One shape: a state plus a set of named questions. Clients written for that format work unchanged.
How it is scored
Decision models are judged on whether their probabilities can be trusted, not only on accuracy. Public benchmarks use Brier score for yes or no questions, the probability mass on the correct option for choice questions, and expected-score error for scales. A well-calibrated 70% should be right about 70% of the time, which is what lets a team set an action threshold on top of the number.
On the Typed Decisions test split (2,000 decisions), the model card reports 0.812 accuracy for v5.
Serving it
The model runs on the standard /v1/systemone endpoint of Hugging Face transformers serve (main branch, October 2026):
hf download FINAL-Bench/Darwin-27B-ZTC-v5 systemone_config.json --local-dir .
transformers serve FINAL-Bench/Darwin-27B-ZTC-v5 --decision-config systemone_config.json
BF16 weights need about 60 GB of GPU memory.
Where it fits
Routing, approval gates, triage, risk checks and agent step guards: any place where a system needs a calibrated yes, no, pick or score, many times a minute, rather than a paragraph.