Skip to main content

The two-stage gate

A consequential action in System9 crosses two independent gates, and both must pass. They ask different questions about different subjects, which is why neither substitutes for the other.

  1. Policy authorization, default-deny, over the actor: may this principal do this at all?
  2. Net potential gain, over the entities the action would affect: is the effect net-positive across everything it touches?

The second is the distinctive one. It evaluates a proposed action by its net potential gain across the entities it would affect, not by whether the actor holds a permission bit. It is an authorization decision made at the decision boundary, not a content filter on a model's input or output, and it can refuse the action outright.

The ordering matters: a principal can be fully entitled and still be refused, because entitlement is a statement about the actor and net gain is a statement about everyone else. A negative net potential gain refuses. A missing input returns insufficient_data, a first-class verdict that never silently defaults permissive, so inability to judge is recorded as inability to judge rather than resolved as consent. The same gate runs on the cell and the Edge Agent, pinned to one contract by cross-language parity fixtures.

One canonical page

The full treatment, the four verdicts, the evaluation algorithm, the tool-audit vocabulary, and the two-runtime parity mechanism, lives under Alignment, where the gate is defined:

→ Agent Platform → Alignment → The NPG gate

Design status

The NPG gate is described by mechanism. Its decisions are designed to be captured as records in a hash-chained audit ledger; the ledger is fed by its callers rather than guaranteeing a row for every decision in the platform. See Compliance & Audit.