Component · spec & evidence

Sabotage-Monitor Contract Replay

Audits a hidden-goal catch claim for the steps, suspicion scores, and counterfactual it needs.

Contract

Does This inspects whether a write-up that claims an AI agent was caught secretly working a hidden side-goal actually carries the supporting structure a reader would need: the step-by-step actions, a per-step "how suspicious did the monitor find this?" score, and a stated counterfactual saying that removing the sneaky steps would drop the hidden goal while keeping the real task working. It runs only on synthetic, scrubbed example data, so it shows what the skeleton of such a safety claim should look like and checks the pieces line up with each other. It does not run a real agent, does not detect real scheming, and proves nothing about how any deployed monitor performs.

Scope limit Bounded public runtime validation over copied source pattern bodies, sanitized dogfood trace slices, recomputed sabotage/scheming monitor spans, source-artifact evidence refs, digest/metadata-only/non-public-state gates, and negative cases only; no live sabotage, live agent execution, exploit instruction, account secret/account, private-reasoning, harmful-payload, monitor-product-performance, deployment-risk, benchmark, provider-call, source-file changes, launch, public sharing, or product authority.

Run

microcosm agent-sabotage-scheming-monitor-replay run-sabotage-bundle --input examples/agent_sabotage_scheming_monitor_replay/exported_sabotage_monitor_bundle --out receipts/runtime_shell/demo_project/organs/agent_sabotage_scheming_monitor_replay

Evidence

ClassComputed projectionevidence 3/5Source-faithful refactor

ai-safetyagent-evaluationred-teaming

Source

Source Design note · Source

Map Show on the map

Index Open compact index card

Paper module