Agi Jev Containment

No. Levels only escalate: L1-L2 record, L3 tags and continues, L4 contains one agent and voice-pages on-call, and L5 environment cut waits for a human. Host scripts exist but never auto-run.

The unit is the run chain, so curl plus reading env plus posting contents reads as intent where one call is noise. Jev answers typed criticality and intent questions over tape plus key nodes.

Yes. Empty env is valid and classification returns degraded instead of failing, while events still land on tape and graph. Live classification needs Jev and supervisor keys plus Docker Compose v2.

Normalize and redact, append JSONL tape, run SafetyDrift and policy, ask Jev, gate the level, then persist to Neo4j and stream over SSE. Agent sandboxes sit on an isolated network away from the dashboard.

Docker Compose v2 and GNU make, and everything runs in containers. Copy .env_template to .env, which is valid empty, then make build to start Neo4j, the API and the UI. The dashboard is on port 3000 with /, /live, /ladder and /trace?run=<id>; the API is on 8000 with /docs and /health; the Neo4j browser is on 7474, logging in as neo4j with password hackspain-local unless you changed the NEO4J_* values. make up, down, restart, logs, backend-shell, frontend-shell, neo4j-shell, lint, test, test-agent and clean cover the rest.

A recorded malicious-agent chain replayed through the whole pipeline, tape to classification to gate to playbooks to Neo4j to SSE, with no live agent and no credentials needed. There are four chains, and every even Trigger is the L5 lateral_db walk while odd Triggers shuffle between exfil, forge and lateral. View trace then opens the per-event timeline for that run.

Yes, by POSTing a JSON event body to /api/runs/{run_id}/events, for example a file_read event with a path. Without TYPESAFE_API_KEY the verdict comes back with degraded: true rather than a 500, and the event still lands on the tape and the graph. Live classification needs TYPESAFE_API_KEY for Jev plus HELMCODE_API_KEY for the supervisor, and running a real sandboxed agent additionally needs make agents-build and make collect.