model · available
Decider 1.1
Give it a piece of state — a support conversation, an alert, an invoice, an agent trace, a screenshot or a receipt — and a set of typed questions about it. Every question comes back as a probability distribution: yes or no, a choice between up to 255 labels, or a score on a rubric. Each question is read on its own, so adding one never moves the answers to the others, and it says "I don't know" when you offer that and the state does not support an answer. It speaks the System One schema, so the typesafe-sdk works against it by changing its base URL.
spec
- model
- decider-1
- context
- 32,768 tokens
- max output
- none — nothing is generated
- input price
- $0.04 / 1M tokens
- output price
- $0.00 / 1M tokens
- repeatability
- same input, same output
- max request
- 32,768 tokens, state and questions
built for
- Routing and triage: which team, which queue, whether a human needs to see it.
- Guardrails and checks inside agents: is this action safe, did this step succeed, should the run stop.
- Scoring against a rubric — severity, urgency, risk — where you want the spread, not just a number.
not for
- Open-ended answers. It picks between the options you give it and cannot write anything else.
- Facts it would need to look up. It decides from the state you send and knows nothing beyond it.
- Locating things in an image (bounding boxes) or reading out text it should copy. It chooses between the options you give.
calling it
One state, every question.
curl https://meragpt.com/v1/systemone \
-H "Authorization: Bearer $MERAGPT_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model": "decider-1", "state": {...}, "questions": {"refund": {"type": "noul", "instructions": "..."}}}'Send one state — any JSON — and as many typed questions as you need: noul for yes or no, choice for up to 255 labels, score for an ordered rubric. Each comes back as a distribution: the probability of yes, a probability per label, or a probability per level with the expected score. All the questions are answered together, so adding one costs a few dozen tokens, and output is never billed. It speaks the System One schema: point the typesafe-sdk at https://meragpt.com with TYPESAFE_BASE_URL and your meraGPT key, and your existing code works. See the docs.
evaluation
How it measures up
Decider 1.1 against TypeSafe Jev 1.14, both run by us on the same items (Jev through its own API). The typed-decisions benchmark has four workflows the model never saw in training — customer service, security incidents, invoice processing and agent traces — with 2,000 decisions scored against a reference distribution.
| typed decisions, accuracy | 0.746 | Jev 1.14 scores 0.726 on the same cases |
| typed decisions, KL from reference | 0.116 | lower is better; Jev 1.14 is 1.471 |
| choices with 33 to 64 options | 0.969 | Jev 1.14 0.887 |
| choices with 65 to 128 options | 0.944 | Jev 1.14 0.906 |
| When2Call (call, ask or refuse) | 77.9 | Decider 1 scored 42.0; Jev 1.14 85.8 |
| WildGuardTest (mean F1) | 91.0 | Decider 1 scored 74.5; Jev 1.14 93.9 |
| speed, per request | 45,000 tokens/s | input read per request on our highly optimised inference; a typical decision takes about 20 ms on the model |
disclosure
Decider 1.1 is our own model, built and trained in house for this one task. The weights are proprietary and are not published. What is published is the measurement: every number on this page comes from a test run on data the model had never seen in training, not from a demo we liked. decider-1 always serves the current version of Decider 1.1: when an improved version ships, every request moves to it, and this page shows its numbers.
Calling it from your own code is in the API docs. Pricing and the credit model are on the homepage.