About this article

• Audience: Operators running soak batches, harness validation, and Phantom gates for transit safety workloads
• Goal: After reading, you can select scenarios from the mass-transit matrix, run batch soaks, and interpret PASS, INCOMPLETE, and FAIL against precomp authority
• Type: How-to

Summary

Ghost Protocol validates transit safety workloads through a scenario matrix exercised by the v3 orchestrator, live MCP harness, and Phantom readiness modules. RF-DETR precomp remains authoritative—cloud escalation applies only on justified clips. Use this guide to pick scenarios for crowding, falls, unattended objects, and fleet self-heal without guessing expected outcomes.

Prerequisites

• Phantom proxy: export PHANTOM_PROXY=https://<phantom-proxy>:8788
• SSH to shadows with YOUR_DEVICE_PASSWORD
• Repo at [repo-root] with tools/v3_shadow_orchestrator.py
• API quick reference for pre-soak health probes

1. Scenario matrix overview

Nine core scenarios cover baseline operation, failure injection, recovery, and advanced modules:

Scenario	Purpose	Typical outcome
normal_short / normal	Baseline live operation	PASS — chunks growing, temp safe, track_id present
recovery_after_stall	Post-encoder-stall recovery	PASS after restart + prune + re-probe
encoder_stall_sim	Simulate encoder silent streak	FAIL/INCOMPLETE when stall rate exceeds soak threshold
disk_critical	Storage CRITICAL injection	FAIL pre-prune; improves after KEEP=30 prune
tracker_enabled	Layer2 tracker + host assigner	PASS — track_id in MQTT ~9–12 Hz
contract-fail-lowhz	Low MQTT contract rate	FAIL contract verify
playback + justified	Playback-ready + MP4 pull	PASS — justified 60s chunks, linear PTS
layer3_depth	Monocular depth + spatial fusion	PASS when depth model active and x/y/z present
video_understanding	Gemini + Video Intel on clips	PASS when proxy cloud modules succeed
hermes-mesh	Peer discovery and cross-A2A	PASS — mesh active, peer ack after activation

Density variants (low/med/high crowding) map to the same gates with different actor briefs—precomp count accuracy must stay at or above 80% on harness repeats.

2. Pre-flight health checks

Before a batch, confirm storage and encoder state—use --prune-first when storage_health is CRITICAL:

curl -s "$PHANTOM_PROXY/fleet/health" | jq .
curl -s -X POST "$PHANTOM_PROXY/tools/call?target=<device-ip>" \
  -H "Content-Type: application/json" \
  -d '{"tool":"get_recording_status","arguments":{}}' | jq '.storage_health, .recent_chunks'

Prune when needed, then restart the app via proxy or oakctl before stall or recovery scenarios.

3. Run a batch soak

Start with dry-run to validate scenario names and matrix wiring:

cd [repo-root]
python3 tools/v3_shadow_orchestrator.py \
  --batch --matrix --dry-run \
  --scenarios normal_short,recovery_after_stall,encoder_stall_sim,disk_critical \
  --visual --update-handoff

Run the live batch with TAP and prune-first:

python3 tools/v3_shadow_orchestrator.py \
  --shadow <device-ip> \
  --batch \
  --scenarios normal_short,recovery_after_stall,encoder_stall_sim,disk_critical,video_understanding \
  --prune-first --visual --tap --update-handoff --hermes \
  --batch-id mass-transit-$(date +%s)

For layer3 promote gates, add --layer layer3_depth and include promote_gate in --scenarios. Video understanding requires justified clips and proxy cloud credentials—see Video understanding.

4. Live MCP harness

Run the in-container harness for precomp authority validation across all scenarios:

# On device via MCP or orchestrator invoke
python3 /app/test_harness.py --mode live-mcp --scenario all --repeats 3

Target at or above 80% count accuracy on RF-DETR precomp across the nine scenarios. Harness complements orchestrator soaks—it does not replace proxy TAP or visual checkpoints.

5. Interpret results and grade

Result	Meaning
PASS	Encoder and thermal samples within thresholds; justified chunks present
INCOMPLETE	Soak ended without final gate marker; probes may still pass
FAIL	Encoder silent or failure rate above 25%; contract or storage blocking

Per-run artifacts under test-results/v3-runs/:

Artifact	Contents
tap.log	Proxy calls, MQTT steps, watcher events
checkpoints/*.jpg + *.json	Snapshots and metrics at phase boundaries
summary.json	Scenario verdict with temp, hz, storage
SMOKE_REPORT	Consolidated gate narrative

Production letter grades use bash cost-sentinel/scripts/grade-production-soak.sh (A ≥ 90, production ≥ 85 with zero critical findings). Phantom playbooks h265_stall and disk_critical trigger self-heal when stall or disk scenarios fail—see Phantom Vision.

6. Cross-scenario dependencies

• Prune before stall recovery on tight storage — historical CRITICAL storage blocks encoder recovery scenarios
• Precomp before cloud — run video_understanding only after justified chunks pass
• Hermes before cross-A2A — confirm mesh active on both shadows before federation scenarios
• Layer2 before layer3 — tracker must be stable before depth fusion promote

Next steps

• API quick reference — health and recording probes before batches
• Soaks & automation — full orchestrator flags and matrix modes
• Video understanding — cloud analysis on justified clips
• Troubleshooting — stall, storage, and contract failures

Operator depth

Live fleet state and harness evidence live in private operator handoff (not published) (private).