Monitor Evidence-Boundary Replay

Replays honest and deceptive agent runs and flags any verdict missing its declared backing evidence.

Contract

Does It replays three made-up agent trajectories (a normal one, one trying to hide a failure, and one overstepping a tool's permissions) and checks that every monitor verdict — including a clean "pass" — comes with the result records needed to re-run and verify it. The result records show exactly which checks each verdict depends on, and any verdict missing its backing evidence gets quarantined instead of trusted. It is a demo over fictional data, not a measurement of how well a real monitor catches anything.

Scope limit Bounded public runtime validation over copied source pattern bodies, sanitized dogfood trace slices, recomputed monitor-verdict spans, source-artifact evidence refs, digest/metadata-only/non-public-state gates, and negative cases only; no live agent execution, monitor product performance, control-eval score, safety-validation, benchmark, provider-call, source-file changes, launch, public sharing, or product authority.

Run

Command

microcosm agent-monitor-redteam-falsification-replay run-monitor-bundle --input examples/agent_monitor_redteam_falsification_replay/exported_monitor_redteam_bundle --out receipts/runtime_shell/demo_project/organs/agent_monitor_redteam_falsification_replay

Evidence

ClassComputed projectionevidence 3/5Source-faithful refactor

Links to Sabotage-Monitor Contract Replay

ai-safetyagent-evaluationred-teaming

Source

Source Design note · Source

Map Show on the map

Index Open compact index card

Paper module

Paper module Agent Monitor Redteam Falsification Replay

Synthetic regression drilldown for a monitor/redteam honesty contract; checks falsification handling, negative cases, non-public-state scan, scope limits.

Read the full write-up

Back to Components index → Reference roster →