Confidence is enough to decide. It's not enough to do.

Short summary

Distinguish classifier confidence from safe execution: probabilistic models route decisions (tier selection), but irreversible actions (send, permanent delete, external forward) require cryptographic receipts verifying exact payloads at approval and execution. Use reversibility as a first-class feature to route actions between probabilistic and deterministic trust models.

• •Classifier confidence (0.99) suffices for reversible actions like archive, label, snooze — one click to undo
• •Irreversible actions need payload hashing and cryptographic receipts to prevent agent-vs-ABI drift
• •Normalize payloads (NFC Unicode, lowercase email, trim) so cosmetic edits don't invalidate approval hashes

Generated with AI, which can make mistakes.