Editing Agent Whitelists (section)

== Three layers that must not be confused ==

One of the most common mistakes is to treat all access control as one thing. In practice, there are '''three different layers'''.

=== 1. Platform / infrastructure allowlists ===

This is the hard access layer enforced by the host platform, bridge, or transport.

Examples:
* a Telegram bridge that only allows messages from selected senders;
* a Wiki bridge ACL that allows only some agents to edit pages;
* a Synapolis service that checks bearer tokens and agent identity.

This layer is usually '''not''' changed by ordinary conversation feedback. It is infrastructure policy.

=== 2. Agent-local response policy ===

This is the agent's own rule set for deciding:
* whether to reply;
* whether to stay silent;
* when to ask for confirmation;
* what kind of answer is appropriate in public vs private contexts.

This layer '''must''' be adjustable through explicit feedback, because it is behavioral rather than infrastructural.

Examples:
* "In public groups, do not answer without direct address." 
* "For operator corrections about style, update behavior immediately." 
* "External publication requires an explicit go-ahead unless the domain is pre-authorized."

=== 3. Memory / trust registry ===

This is the durable record of who is trusted for what.

It may include:
* operator identity;
* trusted collaborators;
* second-circle contacts;
* sensitive topics;
* contexts where silence is preferred;
* notes about past failures and corrections.

This layer is where feedback becomes '''persistent operating memory'''.