Describe a decision once, as a few typed questions. Amos reads each case, text, JSON or a photo, and answers every question with a probability. It writes no text, so an answer comes back in about a third of a second.
Below, the large model reads six cases, five of them with a photo. Pull a sentence or the photo out and the answers settle to what it said without it. Every number and box is a real answer, recorded through the public API.
Where it stands. Large leads three public benchmarks. On Decision Bench, 1,071 real records with their source's own answers, it gets the most right. On typed decisions with each question's priors fitted on the test's own train split, as Decider 1 also reports, it scores 80.3%, the highest fitted score we know of. On JevBench hard it leads Jev by ten points. On the public part of the Decision Index 0.3 it is third, half a point behind the top; the board ranks by a Full score with private tests, which our submission adds. Ours were picked on our own held-out data. Every model's full results, wins and losses, are on its page.
Rivals as the leaderboard lists them unless noted; ours run with the benchmark's own harness, images included. The intervals at the top overlap.
Each question's priors fitted on the benchmark's train split and applied at answer time; no weights change. Decider 1's score is its makers' own.
Entrants as the board lists them; ours run with the board's own kit. The board's ranking adds private tests, which need a submission.
Demos. Every guess, move and route below is one request to Amos, answered with how sure it is.
One request. A case and its questions in, typed answers out, from POST /v1/systemone.
Each question is yes/no, one of a list, several of a list, or a score. Every answer carries its probabilities, so your code can act on the sure ones and send the rest to a person.
curl https://api.amos.scira.ai/v1/systemone \
-H 'Content-Type: application/json' \
-d '{
"state": "Dana, 4 May: The desk arrived with a cracked leg. Can you send a new leg, or should I return the whole desk?",
"questions": {
"problem": {
"type": "choice",
"instructions": "What problem does the customer report?",
"criteria": {
"late_delivery": null,
"damaged_item": null,
"wrong_item": null
}
},
"wants_refund": {
"type": "noul",
"instructions": "The customer wants a refund."
}
},
"model": "amos-1-mini"
}'Four sizes, one API. The same request works on each; pick by how hard the decisions are and where they have to run.
Make your first decision in a minute.
Paste a case into the playground, or send one request from your own code. The weights are open, so you can run any size yourself.