DISTRIBUTED SYSTEMS · AUTHORIZATION · AGENTS
Can consent become a type system for agents?
Abstract. PCHP names identity, purpose, exact scope, validity, delivery, and receipt. The hard engineering question is whether those properties can become machine-checkable invariants across tool calls, instead of natural-language promises around them.
How should scope, purpose, expiry, subject, and revocation compose so an agent cannot accidentally widen a grant while planning a multi-tool task?
HARD CONSTRAINTS
- 01A published registry and a token mint must share one vocabulary.
- 02Action tools must fail closed; read tools must not become covert write paths.
- 03Receipts must make a decision explainable without storing the private payload.
- 04Cross-device state cannot be honestly claimed before it has a durable, revocable model.
A FIRST CONTRIBUTION
Build a small, typed grant algebra and a property-based test suite: no composed plan may execute outside the intersection of its active grants.
BUILD IN THE OPEN
github.com/hushh-labs/hushh-research
hushh.ai/discord