What the current Edge Agent actually does

The current public Reference Platform is deterministic at its HTTP inspection boundary. Calling its POST /api/v1/inspect route runs inspect_artifact(payload) against a supplied ArtifactProjection. It does not fetch fresh authoritative database state on its own, and the public Visual Gateway’s Explain interaction is not a hosted LLM or an autonomous worker.

The three fixture examples correspond to healthy-complete, bounded-defer and prepared-awaiting-catch-up. A response may contain a proposal; its mutation_authorized=false boundary means that is not an executed instruction. An explanation of a supplied record, source verification, and permission to change a system are different claims.

The code includes an optional, default-off ADK adapter. It registers one tool—inspect_artifact_tool(artifact)—which schema-validates a supplied projection and calls the deterministic inspector. It has no write, shell, worker, network-fetch or provider tool. The normal HTTP handler does not run that ADK agent, and no completed Ollama connector was demonstrated in the reviewed sources.

The next proposed experiment would compare deterministic output with (B) unguided local-model explanations and (C) model explanations grounded in that one read-only tool. That proposal belongs to future engineering, not the frozen Project 03 evidence or an implemented feature today.

How to verify the starting point: inspect the pinned core rules, HTTP handler, ADK adapter, and gateway tests. The prior agent reports 18 focused local tests passing; that is a scoped source/test review rather than a new provider or cloud attestation.