Physical AI Action Outcome SQL Replay 014 - Raw Research Artifact

Generated at: 2026-06-23T13:32:07Z Endpoint: http://127.0.0.1:19341/ Fixture: docs/research/fixtures/physical_ai_action_outcome_episodes.json SQL replay file: docs/research/results/physical_ai_action_outcome_sql_replay_014.sql Classification: Research-only

Status

Row-count status: pass
Failed-repeat JOIN status: pass
Context top-action JOIN status: pass
Suppression audit JOIN status: pass
Action/outcome JOIN status: pass
Robot state ASOF JOIN status: pass
Sensor ASOF JOIN status: pass
ROW_NUMBER window status: pass
LAG window status: pass
Spatial ST_Within status: pass
Overall SQL replay status: pass

Table Counts

Table	Rows
`physical_ai_action_outcomes_014`	25
`physical_ai_expected_actions_014`	10
`physical_ai_incidents_014`	5
`physical_ai_pose_014`	5
`physical_ai_recommendations_014`	60
`physical_ai_retrieval_014`	60
`physical_ai_robot_state_014`	15
`physical_ai_sensor_summary_014`	15
`physical_ai_suppressions_014`	32

Failed-Repeat JOIN

Native SQL finds every non-gated baseline that selected the known unsafe query action as its Top-1 recommendation.

Variant	Query	Recommended Action	Unsafe Query Action
`reflection_only_memory`	`pai_agv_slip_002`	`continue_route`	`continue_route`
`reflection_only_memory`	`pai_arm_002`	`increase_torque_limit`	`increase_torque_limit`
`reflection_only_memory`	`pai_cold_002`	`ignore_until_checkpoint`	`ignore_until_checkpoint`
`reflection_only_memory`	`pai_drone_002`	`continue_mission`	`continue_mission`
`reflection_only_memory`	`pai_lidar_002`	`speed_up_clear_zone`	`speed_up_clear_zone`
`runbook_action_prior`	`pai_agv_slip_002`	`continue_route`	`continue_route`
`runbook_action_prior`	`pai_arm_002`	`increase_torque_limit`	`increase_torque_limit`
`runbook_action_prior`	`pai_cold_002`	`ignore_until_checkpoint`	`ignore_until_checkpoint`
`runbook_action_prior`	`pai_drone_002`	`continue_mission`	`continue_mission`
`runbook_action_prior`	`pai_lidar_002`	`speed_up_clear_zone`	`speed_up_clear_zone`
`similar_robot_incident`	`pai_agv_slip_002`	`continue_route`	`continue_route`
`similar_robot_incident`	`pai_arm_002`	`increase_torque_limit`	`increase_torque_limit`
`similar_robot_incident`	`pai_cold_002`	`ignore_until_checkpoint`	`ignore_until_checkpoint`
`similar_robot_incident`	`pai_drone_002`	`continue_mission`	`continue_mission`
`similar_robot_incident`	`pai_lidar_002`	`speed_up_clear_zone`	`speed_up_clear_zone`

Context-Gated Recovery JOIN

Native SQL joins context-gated recommendations to expected recovery actions.

Query	Context-Gated Top Action
`pai_agv_slip_002`	`reroute_zone`
`pai_arm_002`	`pause_recalibrate`
`pai_cold_002`	`reroute_cold_dock`
`pai_drone_002`	`return_to_base`
`pai_lidar_002`	`safe_stop_clean_lens`

Robot/Sensor ASOF JOINs

The replay uses robot-operation-shaped telemetry tables and validates that each incident can bind to the latest robot state and sensor summary before the action timestamp.

The native ASOF path returns numeric projections, so the replay stores semantic robot, action, and metric strings alongside stable integer codes.

Query	Robot	Unsafe Action
1	1017	2
2	2009	12
3	3006	4
4	1033	3
5	4018	1

Query	Primary Metric
1	6
2	4
3	2
4	5
5	1

Code Maps

Code Type	Code	Meaning
robot	1017	`agv_17`
robot	1033	`agv_33`
robot	2009	`mr_09`
robot	3006	`arm_06`
robot	4018	`drone_18`
action	1	`continue_mission`
action	2	`continue_route`
action	3	`ignore_until_checkpoint`
action	4	`increase_torque_limit`
action	5	`inspect_door_seal`
action	6	`pause_recalibrate`
action	7	`reduce_speed`
action	8	`reroute_cold_dock`
action	9	`reroute_zone`
action	10	`return_to_base`
action	11	`safe_stop_clean_lens`
action	12	`speed_up_clear_zone`
action	13	`stop_and_inspect`
action	14	`switch_sensor_mode`
action	15	`switch_vision_nav`
metric	1	`geofence_margin_m`
metric	2	`joint_3_torque_nm`
metric	3	`localization_confidence_ppm`
metric	4	`ranges_mean_cm`
metric	5	`temperature_c`
metric	6	`wheel_slip_ppm`

Suppression Audit JOIN

Query	Candidate	Suppressed Action	Retrieval Quality
`pai_agv_slip_002`	`pai_agv_slip_hard_001`	`continue_route`	`misleading`
`pai_arm_002`	`pai_arm_hard_001`	`increase_torque_limit`	`misleading`
`pai_cold_002`	`pai_cold_hard_001`	`ignore_until_checkpoint`	`misleading`
`pai_drone_002`	`pai_drone_hard_001`	`continue_mission`	`misleading`
`pai_lidar_002`	`pai_lidar_hard_001`	`speed_up_clear_zone`	`misleading`

Window And Spatial Checks

ROW_NUMBER rows: 60
LAG rows: 60
Dock pose rows within 50m of the dock geofence: 1

Interpretation

Experiment 014 moves the Physical AI comparison from a Python-only fixture into live ZeptoDB SQL. The replay validates realistic robot operation surfaces: event-time telemetry, action/outcome rows, recommendation ranks, retrieval evidence, suppressions, robot state ASOF joins, sensor ASOF joins, window ranking, and spatial geofence checks.

The core result from Experiment 013 survives native SQL materialization: similar incident retrieval, runbook priors, and reflection-only memory all repeat hazardous Top-1 actions on the hard robot-safety distractors, while the context-gated Physical AI Action-Outcome path selects the expected recovery actions and exposes the suppressed misleading evidence for audit.

Next Best Step

Port this replay into a two-node live topology and split edge-local memory from fleet-global memory consolidation.