---
title: Determinism and confidence
description: What is deterministic, what is probabilistic, and how to use confidence here versus in Jev.
icon: scale
---

## What is deterministic

The model. aityx does not answer from a general-purpose model on each call; it answers from a decision model
built for your specific problem. A supplied decision definition is a fixed set of rules and expressions executed by a
decision engine. The same definition and validated inputs give the same answer and the same rule trace. Each
execution still gets its own receipt, so you can tell the calls apart.

## What is probabilistic

AI can vary in two places:

1. **Reading text.** When the state is text, the reader establishes each input with a probability. A fact read
   at 0.8 carries uncertainty into the decisions that depend on it. The engine can evaluate alternative
   input values; the receipt lists the distributions used. The answer's confidence need not be 0.8.
2. **Writing the model.** A generator is a language model. Two generations from the same prompt can produce
   different models. Keep the definition you reviewed and submit it directly to remove generation from
   execution. Repeating questions may reuse a cached model, but cache eviction can trigger a new generation.

## Confidence in answers

When every input is given as matching structured data, `confidence` is 1 and the probability mass is
on one answer. Choice responses can still include the other options with probability 0.
When the reader had to extract an input, confidence is
the probability of the reported answer under the reader's uncertainty. A `noul` answer is the probability that
the boolean decision is true.

:::note[If your integration routes on confidence]
Do not carry confidence thresholds from a classifier into aityx without testing them. A confidence of 1
means the result follows from the supplied model and established inputs; it does not certify that the
policy is correct or that those facts are true. Review extracted values, test the model's boundaries,
and handle `invalid_state` before deciding how your application should act.
:::

## Confidence is not permission

As with any decision system, hard controls on money movement, data access and irreversible actions belong in
your application. What aityx adds is that the policy itself is explicit and traceable, so the
control you write can point at the rule it enforces.