Agent Decision Receipts

AI agents should not act in the fog.

This repository is a public proof notebook for Agent Decision Receipts. It shows one idea in plain language: when software acts on behalf of people, the action should leave a readable trace.

What is this?

A small field notebook around Agent Decision Receipts. Each receipt records intent, permission, evidence, refusal or approval, and closeout.

Why does it matter?

A software agent that only talks is useful. A software agent that acts creates an authority problem. Authority needs proof.

What does this notebook show?

• A simple public-safe receipt shape
• Three fictional examples (allowed, refused, refused)
• A short casebook for a generic documentation scenario
• Field notes, scope and limits, and design principles
• A local read-only validator

What does this not prove?

This is a teaching pattern, not a product.

This does not prove production readiness, legal compliance, security, or that agents are safe to run without human oversight.

Start with the pattern first, then read the limits.

Start here

1. FIELD_NOTES.md
2. docs/CASEBOOK_V0.md
3. examples/001_allowed_internal_draft.json
4. examples/002_refused_external_publish_missing_approval.json
5. docs/SCOPE_AND_LIMITS.md
6. BOUNDARIES.md