Why we built Simo

The problem

Is the task finished? Which team handles this? Is this action safe? Which button? What is the order ID? Every agent step, ticket, moderation item and test assertion contains questions like these.

The usual answer is to prompt a chat model and parse what it writes. That approach is:

• Slow. A reasoning model writes a few thousand words before every answer. It takes seconds for a one-word decision.
• Expensive. Those words are billed, on every call, for decisions that are mostly routine.
• Brittle. Your code has to parse prose. Formats drift, and parsers break.
• Unquantified. “I’m confident” is not a number. There is no threshold to set.

Five models to do one job

To avoid the essay, teams stitch together specialists: an intent classifier, a sentiment model, a moderation model, an extractor, and a router that decides when to call the big reasoning model. Five services, five calls, five latencies added up.

Simo is all five in one call: intent, sentiment, policy, extraction and the decision to escalate, from one read of the situation.

What we wanted instead

• A model whose output is a probability you can threshold, not text you must interpret.
• Answers in milliseconds, so judgment can sit inside the loop, not beside it.
• Typed answers, so an integer is always an integer and a list always parses.
• The ability to ask many questions about one situation without paying for the read each time.

The result

Simo is System 1.5: a one-pass reflex with calibrated probabilities, plus the typed values acting requires. It is built at Temprl Labs as the judgment layer for software that acts.

Stop parsing essays. Start reading probabilities.