About this article

• Audience: Integrators and operators planning a Ghost Protocol edge deployment
• Goal: Understand what the 1.0.0-rc5 baseline delivers and how release gates are organized
• Type: Concept

Summary

Ghost Protocol 1.0.0-rc5 ("Mass Transit Safety Baseline") is the first production-oriented edge release for sovereign AI cameras on buses and rail. The baseline stacks four independent capabilities—continuous H.265 archival, RF-DETR safety metrics, Gemma reasoning over A2A, and manifest-driven firmware updates—validated on paired fleet devices before fleet-wide promotion.

Prerequisites

• Two Shadow devices at site-role-a (canary) and site-role-b (secondary) on identical firmware
• Network reachability to each device at <device-ip>
• Phantom proxy or local operator tooling for harness and soak workflows
• Familiarity with production grading themes

Four release tiers

Capabilities ship in priority order. Traditional surveillance is never blocked by AI workloads.

Tier	Capability	What “done” means
1 — H.265 archival	Continuous 4K H.265 chunks, independent of AI	New keyframe-aligned chunks on a regular cadence on both devices; thermals within soak limits; storage never blocks writes
2 — RF-DETR precomp	Person counts, density, and safety flags from on-device detection	Precomputed metrics are authoritative in queries and harness runs; nine mass-transit scenarios pass accuracy gates
3 — Gemma + A2A	Conversational safety reasoning over detections	Signed A2A queries return natural language plus precomp counts in under two seconds; MCP tools callable
4 — Firmware manager	Manifest-driven update signaling	Devices poll a central manifest, surface pending updates on health, CLI, MCP, and A2A surfaces, and accept one-command deploy after publish

Design philosophy

1. Traditional surveillance first — H.265 recording runs 24/7 regardless of NPU or query load.
2. Precomp over recounting — RF-DETR fusion supplies counts and density; the language model synthesizes context, it does not re-count people.
3. Zero-trust agents — External A2A traffic uses HMAC signing, rate limits, and input sanitization.
4. Hermes stays internal — Camera-to-camera coordination remains on the fleet bus, not the public internet.
5. Manifest as release gate — A published firmware manifest is the single source of truth; devices report drift like a managed update channel.

Measurable sign-off gates

Operators promote rc5 when both site roles pass all four tiers:

• Tier 1: Verified H.265 chunks under concurrent AI load; encoder verification harness passes; no prolonged recording gaps during soak.
• Tier 2: Harness accuracy at or above the defined threshold on all nine transit scenarios; A2A responses cite precomputed counts.
• Tier 3: Signed queries succeed with precomp-backed answers; firmware and safety status appear in responses; scenario prompts produce conservative, traceable outputs.
• Tier 4: Manifest publish triggers update-available signaling on every surface within one poll cycle; controlled drift test confirms detection and revert.

Live percentages, probe timestamps, and session evidence are maintained in operator handoff—not in public docs.

Validation workflow

1. Bring both devices to identical rc5 stacks and confirm health endpoints.
2. Run continuous soak with MQTT and visual gates per soak grading.
3. Execute the scenario harness in live MCP mode on each device.
4. Publish a manifest build and confirm update signaling (optional drift test).
5. Record results in grading artifacts and compare against release status themes.

Example operator checks (replace placeholders):

./tools/shadow-cli --ip <device-ip> --health
./tools/shadow-cli --ip <device-ip> --check-update
python3 test_harness.py --mode live-mcp --mcp-url http://127.0.0.1:8765 --scenario all

Locked production decisions

• H.265 chunks rotate on a fixed cadence and are never gated on AI availability.
• RF-DETR precomp is the authority for person count, density level, and safety flags.
• Gemma handles reasoning and tool dispatch; every external query is signed and auditable.
• Firmware version is written at startup, polled periodically, and echoed across five operator surfaces.
• Thermal governor, persistent storage, and conservative alerting policies apply on real vehicles.

Next steps

• Release status — current public baseline and verification checklist
• Production grading — what soak gates measure
• Recovery troubleshooting — when a tier stalls during validation
• Contributing release docs — how public pages ship with releases

Operator depth

Wave-by-wave gate maps, harness JSON, scheduler evidence, and promote decisions live in private operator handoff (not published).