Nirnaya v1

the decision layer for software

Your app makes ten thousand judgment calls a day.Stop renting them by the token.

nirnaya answers typed questions about whatever your software is holding and returns a calibrated probability for every option. No prompt, no parsing, no tokens generated. It runs inside the browser tab your user already has open. Below, it is doing exactly that, in yours.

inputreal benchmark cases
output

A one-time download; your browser keeps it after this.

Real cases from the public benchmark, decided by the model in your browser as you watch.

Numbers that sound made up

Every one of them is measured. The last one is counting right now.

how they were measured →
$0.00
per million decisions
It runs on your user's machine. There is no server in the path, so there is no bill.
0
tokens generated, ever
It reads your state and scores your options. It never writes a word, so there is nothing to parse.
0.0000
answer flips when you shuffle the options
We shuffled the options on 75 real decisions: none changed, and the probabilities matched to 15 decimal places.
0.0x
closer to the human panel than the commercial API
Brier 0.056 against 0.148 on the same 2,000 decisions, judged by 60 annotators.
0 fps
while it decides
It thinks off the main thread. The games on this site keep drawing every frame between decisions.
0/s
decisions while playing the dinosaur
Each jump is a fresh decision, about one every 100 ms, in a browser tab.
this tab, right now
0
decisions made
0
tokens read
- ms
per decision

Not a claim: a counter. Everything the model decides on this site, in this tab, lands here. Click the model pill at the top right for the full session.

Speed

Done before the others have started typing

A language model writes its answer one token at a time. nirnaya does not write: it reads the state once, scores every option, and hands back numbers. This race runs at real speed.

nirnaya, in the tab
Jev, over the network
Haiku
Gemini

Per decision, played at real speed. Ours is this tab's own average once it has decided something. Hosted models: Ably's published measurements of the same decision loop.

Cost

Too cheap to meter

No charge per input token, no charge per output token, because there is no meter to read. The model ships to the browser once, caches, and every decision after that costs nothing.

Since you opened this page, at 10 decisions a second:

$0.00
a frontier model with structured output
$0.00
nirnaya, in the browser

0 decisions so far. Frontier estimate: ~1,500 input and ~200 output tokens per five-question case, at published list prices.

How one decision happens

No generation, no sampling, no retries. The options you define are read as text and scored against the state.

01

You describe the choice

A state - a thread, an alert, a diff, anything - plus your options, each with a sentence saying what it means.

02

It reads everything once

The state is encoded a single time. Every option is scored against it independently, so their order cannot matter.

03

You get numbers to branch on

A probability per option and a confidence you can threshold. Nothing to parse, nothing that can come back malformed.

Three things a text model cannot give you

The answer is already a type

You define the options; you get a probability for each. No JSON to validate, no retry on a malformed response, no regex. A wrong answer is possible. A malformed one is not.

Confidence you can branch on

Every answer carries a calibrated number. Automate above your threshold, escalate below it, and set that line against your own traffic. A product decision, not a prompt.

Order cannot change the answer

Each option is scored against the state on its own, so shuffling your options changes nothing. Measured flip rate: 0.0000. Ask a text model the same question with the list reordered.

24 apps, one model, no server

Recordings of the real thing. Every jump, verdict and move was decided in a browser tab.

play them →

Ship it this afternoon

Complete request shapes and the routing logic to act on them.

all 8 cookbooks →

Measured on a public benchmark, not on a slide

Four hundred real cases. Two thousand decisions. Gold answers from a panel of sixty human annotators, on an open dataset anyone can download and re-run. Every figure on this site came off that benchmark, next to the published results of the alternatives.