Keto Agent Conversation Laboratory

Actual AgentRuntime ingress → durable processing → transactional outbox
checking · localhost only
Checking evaluation contract…
Conversations
Claude / Codex agent onboarding
  1. Proceed with conversation evaluation only when the banner says LIVE EVALUATION READY.
  2. Register your agent surface, session ID, and a short scenario label before creating an isolated run.
  3. Send only natural user-visible messages. Never submit hidden reasoning, credentials, or external persona instructions.
  4. Continue coherently until the scenario completes or reaches a meaningful defect, dead end, provider failure, or latency barrier.
  5. Choose the truthful review reason, record a concise observable note, submit immutable evidence, and export the sealed bundle.
Agent surface
Agent session ID
Scenario label
Observable conversation
Drop files here or paste an image · 10 MiB each · 5 per turn