Physical AI Action-Outcome Experiment 015 Edge/Fleet Replay

Date: 2026-06-23 Status: Research complete Classification: Research-only

Goal

Validate a two-node Physical AI memory shape:

• edge-local memory suppresses unsafe robot actions immediately,
• fleet-global memory receives slower consolidated evidence for audit and cross-robot learning.

Experiment 014 proved the single-node native SQL replay. Experiment 015 splits that evidence across two live ZeptoDB HTTP endpoints.

Hypothesis

Physical AI deployments should not wait for fleet/cloud consolidation before blocking a known unsafe action. A robot or edge controller should make the immediate safety decision locally, then send compact decision, retrieval, and suppression evidence to a fleet node for slower audit and learning.

The expected result is:

• edge-local node chooses expected recovery actions,
• edge-local node suppresses the unsafe query action immediately,
• edge-local ASOF joins bind incidents to current robot/sensor state,
• fleet-global node receives delayed consolidated decisions,
• fleet-global audit JOIN identifies the misleading hard distractors,
• fleet-global window functions preserve retrieval ordering and consolidation lag history.

Inputs

• Fixture: docs/research/fixtures/physical_ai_action_outcome_episodes.json
• Baseline harness: docs/research/tools/physical_ai_action_outcome_baseline.py
• Experiment 014 helper: docs/research/tools/physical_ai_sql_replay.py
• Experiment 015 harness: docs/research/tools/physical_ai_edge_fleet_replay.py
• Edge SQL replay: docs/research/results/physical_ai_edge_fleet_replay_015_edge.sql
• Fleet SQL replay: docs/research/results/physical_ai_edge_fleet_replay_015_fleet.sql
• Result: docs/research/results/physical_ai_edge_fleet_replay_015.md

Procedure

1. Start an edge-local ZeptoDB server on port 19441.
2. Start a fleet-global ZeptoDB server on port 19442.
3. Load the Physical AI fixture and recompute the Experiment 013 context-gated comparison.
4. Materialize edge-local tables:
   • incidents,
   • expected actions,
   • robot state,
   • sensor summaries,
   • immediate decisions,
   • immediate suppressions.
5. Materialize fleet-global tables:
   • expected actions,
   • historical action outcomes,
   • delayed edge decisions,
   • top-3 retrieval evidence,
   • suppression audit rows.
6. Validate edge-local SQL:
   • row counts,
   • recovery-action JOIN,
   • risky-action suppression count,
   • robot state ASOF JOIN,
   • sensor ASOF JOIN.
7. Validate fleet-global SQL:
   • row counts,
   • consolidated recovery-action JOIN,
   • delayed consolidation lag,
   • suppression audit JOIN,
   • retrieval ROW_NUMBER,
   • consolidation-lag LAG.

Boundary

This is a research-only harness. The delayed edge-to-fleet transfer is modeled by Python writing SQL rows to two live ZeptoDB endpoints. It is not a new runtime replication, routing, or control-plane feature.

Command

./build/zepto_http_server --port 19441 --node-id 1 --no-auth --storage-mode pure

./build/zepto_http_server --port 19442 --node-id 8 --no-auth --storage-mode pure

python3 docs/research/tools/physical_ai_edge_fleet_replay.py \
  --edge-url http://127.0.0.1:19441/ \
  --fleet-url http://127.0.0.1:19442/ \
  --edge-stats-url http://127.0.0.1:19441/stats \
  --fleet-stats-url http://127.0.0.1:19442/stats \
  --fixture docs/research/fixtures/physical_ai_action_outcome_episodes.json \
  --output docs/research/results/physical_ai_edge_fleet_replay_015.md \
  --edge-sql-output docs/research/results/physical_ai_edge_fleet_replay_015_edge.sql \
  --fleet-sql-output docs/research/results/physical_ai_edge_fleet_replay_015_fleet.sql \
  --timeout 10

Acceptance Criteria

Criterion	Required
Harness compiles with py_compile	pass
Edge server accepts all DDL/DML	pass
Fleet server accepts all DDL/DML	pass
Edge row counts	pass
Fleet row counts	pass
Edge recovery-action JOIN	pass
Edge risky-action suppression	pass
Edge robot ASOF JOIN	pass
Edge sensor ASOF JOIN	pass
Fleet consolidated recovery JOIN	pass
Fleet suppression audit JOIN	pass
Fleet consolidation lag	pass
Fleet ROW_NUMBER	pass
Fleet LAG	pass

Result

See docs/research/results/physical_ai_edge_fleet_replay_015.md.

Summary:

• Overall edge/fleet replay status: pass.
• Edge-local node stored 82 research rows.
• Fleet-global node stored 87 research rows.
• Edge-local node selected all five expected recovery actions before consolidation.
• Edge-local node suppressed all five unsafe query actions immediately.
• Fleet-global node received five delayed decisions with lag >= 250 ms.
• Fleet-global audit JOIN exposed all five misleading hard distractors.
• Edge ASOF, fleet JOIN, fleet ROW_NUMBER, and fleet LAG checks passed.

Interpretation

The experiment validates the product thesis for Physical AI memory separation: the robot/edge path owns immediate safety, while the fleet path owns slower global learning and audit.

The result also sharpens the next engineering question. The memory primitive is useful, but production needs an explicit edge-to-fleet consolidation path with idempotency, ordering, late-arrival handling, and loss/duplicate telemetry.

Next Product Or Research Step

Replace harness-driven consolidation with a bounded edge-to-fleet feed or replication path, then test:

• dropped consolidation rows,
• duplicate consolidation rows,
• late-arriving fleet audit evidence,
• edge operation during fleet node outage,
• fleet replay after restart.

Status update: Experiment 016 completed this research step with an explicit bounded edge outbox, fleet inbox, ACK, telemetry, duplicate, dropped, late, outage, and restart replay harness. The next step is runtime connector promotion, not another direct harness copy.