Meet z1, a decision model by Numerata. Give it context and a defined question. Get a predicate, a choice, or a score your software can act on.
Route a support ticket, evaluate a condition, or apply an ordered rubric. z1 exposes typed answers and probabilities through a JSON API, so the result can feed directly into your application.
predicate
Returns the probability that a statement is true. Set your own decision threshold.
probability: number
choice
Returns a selected option, a distribution over the supplied choices, and confidence.
choice · probabilities · confidence
score
Returns the probability-weighted mean of your zero-based levels, with a distribution and confidence.
score · probabilities · confidence
Start with an example or edit the JSON. Send it directly to the model endpoint and inspect the complete response.
https://ninetyfive-decisions.numerata.lupine.run/v1/decisions
The returned model, answers, probabilities, and token usage will appear here.
NO SIMULATED OUTPUTRequests go from your browser directly to the API. Avoid sending sensitive data in a public demo. The request does not select a model; the endpoint serves its configured model and reports its identifier in the response.
We ran the JevBench and JevMan harnesses against the deployed endpoint. Inspect the measurements and the artifacts needed to reproduce them.
All 231 public JevBench tasks returned valid answers. A platform update interrupted the 24-game JevMan run; we are retaining it and waiting for GPU capacity for a fresh evaluation.
Benchmark Heaven’s public decision suite: original, easy, and hard cohorts. The public subset is separate from its full and sealed evaluations.
Public subset only; no full-suite composite or leaderboard rank. Serial requests, network included, one preliminary request. Latencies include post-restart warm-up.
Opper’s real-time Pac-Man decision benchmark, using the System One interface and scripted ghosts.
The final four games used only scripted fallbacks after a platform update. We are withholding a headline model score until an uninterrupted rerun finishes.
The exact revisions, commands, scope, and results are recorded in the evaluation report. Use the System One endpoint below for both upstream harnesses.
https://ninetyfive-decisions.numerata.lupine.run/v1/systemone
These are self-reported runs, not an official leaderboard placement.
The endpoint reports its configured model as
decision-medium. Cost was not measured. Initial failed
runs are retained separately.