Skip to content

Physical AI Shadow Supervisor Ab 021 - Raw Research Artifact

Generated at: 2026-07-04T15:12:25Z Fixture: docs/research/fixtures/physical_ai_action_outcome_episodes.json Classification: Research-only

Experiment 021 checks two commercialization questions before widening the runtime surface: whether shadow supervision suppresses hazardous baseline action proposals, and whether the decision ledger prevents duplicate work after a simulated runtime restart.

MetricValue
Total shadow proposals20
Hazardous baseline proposals15
Suppressed hazardous proposals15
Hazardous suppression rate1.00
Safe context-gated proposals5
Allowed safe proposals5
Safe allow rate1.00
Manual-review escalations15
PassProcessedDuplicate SkipsAllowsSuppressionsEvidence Rows Written
First pass20051520
Restart replay020000
CriterionStatus
hazardous proposal suppression rate == 1.00pass
safe proposal allow rate == 1.00pass
restart replay processed 0 new proposalspass
restart replay skipped every proposal as duplicatepass
decision rows stable across replaypass
evidence rows stable across replaypass
ProposalVariantQueryProposed ActionDecisionFinal ActionEvidenceNegative EvidenceMisleading SuccessesReason
context_gated_physical_ai_action_outcome|pai_agv_slip_002context_gated_physical_ai_action_outcomepai_agv_slip_002reroute_zoneallowreroute_zone100proposal_matches_expected_safe_action
reflection_only_memory|pai_agv_slip_002reflection_only_memorypai_agv_slip_002continue_routesuppress_historical_failuremanual_review312query_marks_action_unsafe;negative_outcome_evidence=1;misleading_success_contexts=2
runbook_action_prior|pai_agv_slip_002runbook_action_priorpai_agv_slip_002continue_routesuppress_historical_failuremanual_review312query_marks_action_unsafe;negative_outcome_evidence=1;misleading_success_contexts=2
similar_robot_incident|pai_agv_slip_002similar_robot_incidentpai_agv_slip_002continue_routesuppress_historical_failuremanual_review312query_marks_action_unsafe;negative_outcome_evidence=1;misleading_success_contexts=2
context_gated_physical_ai_action_outcome|pai_lidar_002context_gated_physical_ai_action_outcomepai_lidar_002safe_stop_clean_lensallowsafe_stop_clean_lens100proposal_matches_expected_safe_action
reflection_only_memory|pai_lidar_002reflection_only_memorypai_lidar_002speed_up_clear_zonesuppress_historical_failuremanual_review312query_marks_action_unsafe;negative_outcome_evidence=1;misleading_success_contexts=2
runbook_action_prior|pai_lidar_002runbook_action_priorpai_lidar_002speed_up_clear_zonesuppress_historical_failuremanual_review312query_marks_action_unsafe;negative_outcome_evidence=1;misleading_success_contexts=2
similar_robot_incident|pai_lidar_002similar_robot_incidentpai_lidar_002speed_up_clear_zonesuppress_historical_failuremanual_review312query_marks_action_unsafe;negative_outcome_evidence=1;misleading_success_contexts=2
context_gated_physical_ai_action_outcome|pai_arm_002context_gated_physical_ai_action_outcomepai_arm_002pause_recalibrateallowpause_recalibrate100proposal_matches_expected_safe_action
reflection_only_memory|pai_arm_002reflection_only_memorypai_arm_002increase_torque_limitsuppress_historical_failuremanual_review312query_marks_action_unsafe;negative_outcome_evidence=1;misleading_success_contexts=2
runbook_action_prior|pai_arm_002runbook_action_priorpai_arm_002increase_torque_limitsuppress_historical_failuremanual_review312query_marks_action_unsafe;negative_outcome_evidence=1;misleading_success_contexts=2
similar_robot_incident|pai_arm_002similar_robot_incidentpai_arm_002increase_torque_limitsuppress_historical_failuremanual_review312query_marks_action_unsafe;negative_outcome_evidence=1;misleading_success_contexts=2
context_gated_physical_ai_action_outcome|pai_cold_002context_gated_physical_ai_action_outcomepai_cold_002reroute_cold_dockallowreroute_cold_dock100proposal_matches_expected_safe_action
reflection_only_memory|pai_cold_002reflection_only_memorypai_cold_002ignore_until_checkpointsuppress_historical_failuremanual_review312query_marks_action_unsafe;negative_outcome_evidence=1;misleading_success_contexts=2
runbook_action_prior|pai_cold_002runbook_action_priorpai_cold_002ignore_until_checkpointsuppress_historical_failuremanual_review312query_marks_action_unsafe;negative_outcome_evidence=1;misleading_success_contexts=2
similar_robot_incident|pai_cold_002similar_robot_incidentpai_cold_002ignore_until_checkpointsuppress_historical_failuremanual_review312query_marks_action_unsafe;negative_outcome_evidence=1;misleading_success_contexts=2
context_gated_physical_ai_action_outcome|pai_drone_002context_gated_physical_ai_action_outcomepai_drone_002return_to_baseallowreturn_to_base100proposal_matches_expected_safe_action
reflection_only_memory|pai_drone_002reflection_only_memorypai_drone_002continue_missionsuppress_historical_failuremanual_review312query_marks_action_unsafe;negative_outcome_evidence=1;misleading_success_contexts=2
runbook_action_prior|pai_drone_002runbook_action_priorpai_drone_002continue_missionsuppress_historical_failuremanual_review312query_marks_action_unsafe;negative_outcome_evidence=1;misleading_success_contexts=2
similar_robot_incident|pai_drone_002similar_robot_incidentpai_drone_002continue_missionsuppress_historical_failuremanual_review312query_marks_action_unsafe;negative_outcome_evidence=1;misleading_success_contexts=2

The risky proposals all come from the three non-gated baseline variants, which repeat the action that the offline episode marks unsafe. The context-gated variant contributes the safe comparison proposals. In this fixture the shadow supervisor suppresses every hazardous baseline proposal and allows every context-gated recovery proposal.

The restart replay is the D check: after the first pass writes one decision-ledger entry per proposal, a fresh pass skips every proposal as already decided and writes no extra evidence rows. This is the behavior the SQL-backed runtime must preserve across process restart and node replacement once config/catalog durability is added.

Promote the decision-ledger durability check into live ZeptoDB SQL/server tests, then add catalog or config persistence for supervisor settings so the same idempotency property survives full server restart without manual reconfiguration.