September 22, 2026 · 5 min

Decider 1: typed decisions, more accurately than Jev, for less

Most of what software asks a language model is not a request for prose. Is this customer asking for a refund? Which team should take this ticket? Did the agent’s last step succeed? How severe is this alert, on our four-level scale? These are decisions, and the useful answer to each is a probability, not a paragraph.

Decider 1 (sd-1) is our model for exactly that. You send a piece of state and a set of typed questions about it; every question comes back as a distribution, and all of them are answered in one call. On the public typed-decisions benchmark it is the most accurate model we know of, ahead of TypeSafe’s Jev, and it costs less per token.

Three kinds of question

Decider 1 speaks the System One schema, which has three question types:

• noul, a yes-or-no question. The answer is the probability of yes.
• choice, a pick between up to ten labels you describe. The answer is a probability for each label.
• score, a position on an ordered rubric you write. The answer is a probability for each level, and the expected level.

Because the answer is read from the model’s scores for the options you offered, it can only ever be one of them: there is no free text to parse and nothing to hallucinate outside the schema. Nothing is generated, so there is no output to bill, and asking twenty questions about one state costs about what one costs, because the state is read once.

The results

The typed-decisions benchmark has 400 cases across four workflows — customer service, security incidents, invoice processing and agent-trace observability — with five questions each, 2,000 decisions in all. Decider 1 had never seen any of the four workflows or any of the twenty question schemas: every number below is zero-shot. We ran TypeSafe’s Jev 1.13.0 through its own API on the same cases.

The fitted baselines were trained on each workflow’s own labels, an advantage Decider 1 did not get. It beats Jev on every question type:

Accuracy is only half of it. The KL column measures how far each model’s probabilities sit from the reference distributions, and there the gap is fifteenfold: when Decider 1 says 70%, it tends to mean 70%. That matters more than the top answer when your code decides what to do with a confidence — to act, to ask a human, or to wait.

How to read these numbers

The reference answers on this benchmark are the average of three samples from a set of teacher models, so a score measures agreement with that ensemble rather than ground truth. Two single teacher samples agree with each other 73.5% of the time. Decider 1 scores above that, which sounds impossible and is not: a model that predicts the ensemble’s average can match the reference better than any one draw from it does. Every figure here is from a single evaluation run.

Speed

Measured end to end from a client, network included: a single request takes about half a second at the median, answering every question in it. A single serving worker sustains around 14,000 prompt tokens a second — about 80 decisions a second on these cases — before requests start to queue.

Switching from Jev

If you already use the typesafe-sdk, point it at us and use your meraGPT key. Nothing else in your code changes:

export TYPESAFE_BASE_URL=https://meragpt.com
export TYPESAFE_API_KEY=$MERAGPT_API_KEY

Two differences are worth checking first. A request — state and questions together — can hold up to 4,096 tokens, which covers typical tickets, alerts and traces but not whole documents. And a choice offers at most ten labels; for more, split the set into groups with a “none of these” label and ask in two steps.

Pricing is $0.03 per million input tokens, against Jev’s $0.042, and output is free on both. There is no subscription: you buy credit and it is drawn down per token.

Try it

The playground runs Decider 1 without an account: paste a state, write your questions, and see the distributions. The /v1/systemone docs have the full request and response shapes, and the model page has the benchmarks and the limits.