
SYGNET — a future policy and evidence control plane for AI agents.
This page describes an architecture preview only. Every element below is a design intent positioned after ratified M1 completion, separated from the current SBBX product scope.
FLOW
Signals → Policy evaluation → Human approval → Constrained action → SBBX evidence
Authority moves in one direction only: signals propose, humans authorise, evidence records.
STAGE EXPLORER
Examine the proposed control path
Conceptual detail only. Each stage remains part of the post-M1 architecture preview.

Proposed 3D stage illustration · explanatory only
Signals are the proposed entry point for observations, requests and events. They provide information to evaluate; they do not grant permission to act.
How it is intended to work
Each input would carry its source, intended purpose and relevant context. The design calls for checking its format, origin and permitted scope before it can become a proposed action. Instructions embedded in external content would remain input data, rather than becoming new authority.
Example
A monitored source reports a policy change. SYGNET could prepare a proposal for review, with a reference to the source and an explanation of what may need attention.
Boundary
Receiving a signal would not authorise an agent to change a system, disclose information or widen its access.
Conceptual supporting diagram · explanatory only

Proposed 3D stage illustration · explanatory only
The policy stage is intended to compare a proposed action with explicit rules: who may request it, which resources it concerns and what limits must apply.
How it is intended to work
Evaluation would consider the action, its target, the requested permissions and relevant conditions. The result should explain whether the proposal fits the allowed scope, must be refused or needs a human decision. Missing authority should prevent execution.
Example
A proposal may permit preparing a document for review while withholding permission to publish it or send it outside the organisation.
Boundary
A favourable policy evaluation would not replace human approval where approval is required, and it would not create permissions that were never granted.
Conceptual supporting diagram · explanatory only

Proposed 3D stage illustration · explanatory only
A person should be able to understand and decide on the specific proposed action before granting approval.
How it is intended to work
The approval view would show the target, intended effect, requested access and meaningful limits. A person could approve that bounded proposal or refuse it. A material change to the proposal would require a fresh decision rather than silently inheriting earlier approval.
Example
A reviewer approves preparing a report for a named internal audience. That decision would not also authorise public publication or unrelated follow-up actions.
Boundary
Silence would not count as approval. An agent’s recommendation would not substitute for the person’s decision.
Conceptual supporting diagram · explanatory only

Proposed 3D stage illustration · explanatory only
Execution is intended to remain inside the permissions and limits attached to the approved action.
How it is intended to work
The action would be checked against its authorised target and allowed operations. Resource, time or usage limits could further constrain execution. A request outside those boundaries should stop rather than silently receive broader privileges.
Example
An approved task could prepare a draft in a designated workspace while remaining unable to publish it, delete unrelated files or access other workspaces.
Boundary
Approval for one task would not become unrestricted autonomy. Changed conditions or insufficient authority should require a new decision.
Conceptual supporting diagram · explanatory only

Proposed 3D stage illustration · explanatory only
The proposed evidence stage would connect the originating signal, the policy decision, the human authorisation and the reported action outcome into a reviewable record.
How it is intended to work
Records would distinguish what was requested, what was approved, what was attempted and what was actually observed. References to the applicable policy and relevant evidence would help a reviewer examine that sequence. Missing evidence and unresolved outcomes should remain explicit.
Example
A reviewer could trace a completed draft back to the request and approval that permitted its creation, while seeing that publication was outside the approved scope.
Boundary
A recorded event would not by itself prove that an action was correct, lawful or complete. Evidence integrity and the truth of the underlying event are separate questions.
Conceptual supporting diagram · explanatory only