Metadata-Version: 2.4
Name: jev-typed-decision
Version: 0.1.0
Summary: Typed decision records for small local models: parse, verify and route confidence instead of prose. Reference notes: https://jev001.org/
Author: Jev001 maintainers
License: MIT
Project-URL: Homepage, https://jev001.org/
Project-URL: Documentation, https://jev001.org/what-is-jev/
Project-URL: Guide, https://jev001.org/jev-api/
Project-URL: Examples, https://jev001.org/jev-use-cases/
Keywords: llm,decision,confidence,routing,typed-output
Classifier: Programming Language :: Python :: 3
Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
Requires-Python: >=3.9
Description-Content-Type: text/markdown

# jev-typed-decision

Small models write convincing prose. Convincing prose is not a confidence
signal. This package keeps a **typed decision record** instead: the answer, a
numeric confidence, the evidence it leaned on, and what to do when confidence
is low.

```python
from jev_typed_decision import record, needs_human

rec = record("route to human", 0.42, evidence=["eval-2026-09"])
needs_human(rec)   # True -> escalate instead of acting
```

## Why

Reading the *shape* of a decision beats reading the *sound* of the answer.
Details, background and worked examples:

- [What Jev is](https://jev001.org/what-is-jev/)
- [Jev API notes](https://jev001.org/jev-api/)
- [Run Laya locally](https://jev001.org/run-laya-locally/)
- [Use cases](https://jev001.org/jev-use-cases/)
- [Home](https://jev001.org/)

## Scope

Reference helper, stdlib only, no runtime dependencies. Confidence clamping is
deliberately blunt; calibrate per project.
