How Simo works

The three steps

1. Show it a situation. Text, a screenshot, or video. Simo reads it once.
2. Ask typed questions. Yes/No, choose one, rate on a scale, fill a value, extract from the input. Ask ten in one request; each adds only milliseconds.
3. Read the probabilities. Every decision is a number you can put a threshold on. Where acting needs more than a verdict, Simo supplies the value too.

How an answer is created

A chat model produces an answer by writing words, one after another, and your code has to parse the result. Simo does not write words to decide. For a Yes/No, Choose-one or Rate question, the answer is the probability: no tokens are sampled to reach a decision.

Because every question is typed, the answer arrives in a shape your code already understands. A Choose-one question returns a probability for each option. A Yes/No returns the probability of each side. A scale returns a value on the scale you defined.

Is the task finished?            →  no: 0.93, yes: 0.07
What should the agent do next?   →  press: 0.89, scroll: 0.06, stop: 0.05
How frustrated is the customer?  →  2.4 of 3

Why ten questions cost little more than one

Simo reads the situation once, whether it is a document, a screenshot or a video. Every question then runs as a cheap branch off that single read. Reading is the cost; questions are nearly free.

On Simo-1, one question over a 1280×720 screenshot takes 64 ms and ten take 104 ms, about 10 ms for each added question. See the full latency numbers.

When acting needs a value, not just a verdict

Deciding to press a button is not enough; the agent needs to know which one. For these cases Simo can fill a value (the button, the amount to scroll) or extract from the input (the order ID, the invoice amount). Filled values are schema-constrained, so an integer is always an integer and a list always parses, and each produced value can carry its own calibrated self-check.

Fill and Extract are in preview. Read more on the question types page.

What Simo never does

• It never writes an essay, a chain of thought or a paragraph of reasoning before answering.
• It never returns an unquantified “I’m confident”. Every decision comes with a number.
• It is not the controller. It supervises the agent, the pipeline or the test suite; it does not drive motors or game input.