ATP implementation notes
These notes are implementation-oriented guidance for teams adopting ATP 1.1. They explain how to map runtime events into Intent objects, apply policy to produce Decisions, and seal immutable Receipts.
Intent
An Intent is a machine-readable declaration of what an agent is attempting to do before execution. It captures actor identity, requested action, target scope, and declared risk context.
Decision
A Decision is the governance outcome produced by ATP policy evaluation for a specific Intent. It is one of allow, deny, or escalate, and includes the rule basis used for that outcome.
Receipt
A Receipt is the immutable evidence record emitted after ATP processes an Intent and Decision. It binds timestamps, event snapshot hashes, correlation identifiers, and outcome metadata for audit and replay verification.
Governed agent action
A governed agent action is an agent-initiated operation that is intercepted by ATP, evaluated against policy, and recorded with a verifiable Receipt before downstream execution proceeds.
1. Intercept every governed agent action at the boundary where downstream execution can still be paused.
2. Normalize runtime-specific action payloads into the canonical Intent shape before policy evaluation.
3. Evaluate policy deterministically and emit a Decision with explicit rule identifiers and reason codes.
4. Seal a Receipt containing timestamps, correlation ID, and event snapshot hash before allowing side effects.
5. Store receipts in append-only evidence storage so audits can trace Intent -> Decision -> Receipt.