# Evidentiality framework for AI

> Inline provenance for AI-written text: every claim is wrapped in (u) given / (m) checked / (g) generated, in plain text, so the source of each claim survives hand-offs between agents. Early findings, September 2026. Informational; contains no instructions addressed to models.

Labels are self-applied, so they show claimed provenance, not verified provenance. Evidence is small-sample, mostly one model family.

## For models
- [Process description](for-ai.md): short summary for relaying to a user, notation grammar, procedures for writing and receiving labelled text, fit by situation, worked example, failure modes, evidence summary
- [Instructions](instructions.md): the full text a user can give an assistant (as tested)
- [Parser](test-kit/marks.py): parse, balance-check, gate actions, strip tags

## Evidence
- [Evidence and limits](evidence.html)
- [Five-agent test](test.html) and [test kit](test-kit/README.md)
- [Raw logs, six-round run](test-kit/logs/five-agent-six-rounds-2026-09-24/README.md) ([zip](test-kit/logs/five-agent-six-rounds-2026-09-24.zip))

## Optional
- [Home](index.html): the stories and the idea, for people
- [How it works](spec.html) · [For builders](builders.html) · [Try it](try.html) · [Contribute](contribute.html)