Appearance
About this article
- Audience: Contributors who need context on retired features and documentation reorganizations
- Goal: After reading, you understand why Gemma NPU reasoning was retired and how phase pivots shaped today's sovereign + precomp stack
- Type: Concept
Summary
Ghost Protocol moved from large on-device language models to RF-DETR precomp authority with proxy-mediated cloud escalation. Earlier phase-one mobile and actor testing plans were archived; current work centers on sovereign edge app bundles, Phantom federation, and stacked release gates.
Prerequisites
- Glossary — terms such as precomp, sovereign, and justified recording
- Architecture overview — present-day system map
Gemma NPU retired (2026)
Before the pivot, edge devices could host a large Gemma-class model on the NPU for primary reasoning. That path added image size, thermal load, and operational complexity without improving the authoritative detection contract.
Current model:
- On-device: RF-DETR precomp from ghost_core; fusion and recording in ghost_agent_v2
- Cloud (optional): Gemini File API and Video Intelligence on justified ~60 s clips only
- Sovereign rule: No cloud API keys baked into the edge app; the Phantom proxy owns escalation credentials
Health and A2A surfaces may still note graceful NPU-offline messaging while precomp remains the trusted detection source. The Ghost Protocol custom-trained artifact name remains ghost_protocol_gemma4_e2b.litertlm—listed with RF-DETR fleet packages on the RF-DETR ML Pipeline hub. Historical Gemma planning documents live under appendices/internal-plans/ and do not sync to the public site.
Phase pivots and documentation reorg
| Phase | Focus | Status |
|---|---|---|
| Phase 1 (early) | Mobile testing, actor scenarios, inter-device kanban, VMS integration drafts, HITL experiments | Archived to internal plans |
| Phase 2 (current) | Precomp + H.265 primary + sovereign deploy; layer 3 depth; Hermes mesh; Phantom readiness; proxy MCP federation | Active public docs subject |
Reorg principles (still in effect):
- Move historical PRDs and checklists with
mvonly—no deletes from the tree - Keep user-facing categories (quickstart, guides, reference, architecture, troubleshooting) free of verbatim handoff dumps
- Index and sidebar list appendices separately from internal-only material
What changed in verification culture
Early work emphasized exploratory mobile and actor flows. Production verification now standardizes on orchestrator batches, soak gates, harness scenarios, and Phantom module readiness before fleet-wide promote. Exact pass rates, soak tables, and session chronology stay in private operator handoff—not in public history pages.
Next steps
- Release gates — current T1–T4 public summary
- Checklists — operational lookup after deploy
- Contributing to docs — how internal plans stay out of public sync
Operator depth
Full chronology, moved-file inventories, and live remainder tracking live in private operator handoff (not published) (private).