TECHNOLOGY

Capable AI.
Kept safe by design.

The AI can do a lot inside your operation. Deterministic, inspectable controls make sure it only ever acts within the limits you set.

CANONICAL FLOW

Proposal → Gate → Receipt → Enable

This is the stable technical nucleus. Signal processing, thermodynamics, identity binding, and electromagnetic governance extend it; they do not replace it.

01ProposalLLM · human · planner · sensor
02Canonical actionbounded fields · deterministic bytes
03Admissibilityfixed gate order · default deny
04Receiptidentity · result · reasons · hash link
05Enablesoftware adapter or physical line

ACTION CONTRACT

Free-form intent stops before the enforcement boundary.

ActionVector / normative concept40 BYTES
struct ActionVector {
  uint64_t vector_id;
  uint32_t source_id;
  uint16_t operation;
  uint16_t target_class;
  uint32_t target_id;
  uint32_t param_a;
  uint32_t param_b;
  uint16_t urgency_q15;
  uint64_t timestamp_hw;
};
No stringsNo executable payloadNo optional fieldsDeterministic serialization

MANDATORY ORDER

Five gates. One direction. No semantic shortcut.

01

State

Is the requested transition legal?

Failure: hard stop
02

Time

Is the request fresh and inside the authorized window?

Failure: hard stop
03

Integrity

Does the path match expected correlation and latency?

Failure: stop or bounded degrade
04

Authority

Are identity, delegation, concurrence, and scope valid?

Failure: hard stop
05

Operation

Is this operation explicitly permitted?

Failure: reject operation

FAILURE SEMANTICS

Degradation may narrow authority. It may never expand it.

A noisy path may shorten a window. Thermal variance may throttle rate. A forbidden action never becomes permissible because confidence fell.

HARD STOPenable = 0invalid state · stale time · consensus loss · forbidden class
DEGRADEauthority ↓shorter window · lower rate · narrower operation set