About this article

• Audience: Operators, integrators, and compliance reviewers who need RF-DETR contracts without reading firmware source
• Goal: Look up model identity, performance expectations, precomp authority rules, and integration hooks
• Type: Reference

Summary

Ghost Protocol ships RF-DETR small as the on-device precomp model for person and weapon detection. Person class c=0 with confidence ≥ 0.35 is the authoritative safety count—used in MCP tools, MQTT, ONVIF events, and soak gates before any cloud escalation. This card documents the production precomp contract.

For the full story—custom Omni packages (rf_detr_oak4.tar / rf_detr_qcs8550.dlc), TrOCR encoder/decoder cascade, ML Pipeline training and turnaround, dataset types, S1/S2/VMS deploy, USA hosting posture, and exact names including Gemma 4—see RF-DETR, TrOCR, ML Pipeline & fleet deployment.

Prerequisites

• Edge AI pipeline — dual-stream context
• ONVIF reference — person events bridged from RF-DETR
• Proxy reachable for status probes — API quick reference

Model details

Field	Value
Name	RF-DETR small (fixed inference build)
Artifact	rfdetr_small_inference_model_fixed.dlc (~135 MB float16)
Input	640×640 NHWC BGR888i
Output	Up to 300 detections with class and confidence
Classes	23 labels — person (0) plus weapon categories
Target	Edge AI / SNPE on Hexagon DSP (RVC4 Shadow 1)
Load path	Bundled under /app/models or /data/models via layer env

Supporting files: labels.json, config.json, buildinfo.json in the models directory.

Intended use

• Sovereign on-device detection for transit-safety surveillance (bus, train, shuttle)
• ~10 Hz inference stride (RF_DETR_STRIDE_FPS=10) on the detection substream
• Sacred authority for person counts—no LLM override
• Upstream input to ONVIF person events, Hermes enrichment, and cloud escalation on justified clips

Out of scope: Primary face recognition on RF-DETR alone. Optional face identity on S2 (Shadow 2) uses person c=0 head crops via nist_fr (edge FATE depth PAD + quality) + host AdaFace IR-18 (512-dim, MIT) gallery match with FRTE 1:1 FMR-calibrated τ (self-assessed, not NIST-certified). S1 (Shadow 1) has no stereo depth — nist_fr does not run there. See NIST compliance.

Key parameters

PERSON_CLASS_ID = 0
WEAPON_CLASS_IDS = {10, 11, 12, 13}  # gun, knife, rifle, shotgun

def count_persons(dets, conf_thresh=0.35) -> int:
    # Count detections where c == 0 and p >= conf_thresh
    ...

Parameter	Default	Usage
PERSON_CLASS_ID	0	Person label in labels.json and all public integrations
conf_thresh	0.35	Precomp person count gate
RF_DETR_STRIDE_FPS	10	Layer environment stride target

Precomp metrics inject first in every A2A, MCP, and Hermes safety query. Downstream LLM layers describe and contextualize—they do not replace counts.

Detection flow

Probe live status:

curl -s -X POST "$PHANTOM_PROXY/tools/call?target=<device-ip>" \
  -H "Content-Type: application/json" \
  -d '{"tool":"get_rfdetr_status","arguments":{}}'

Performance and gates

Check	Expectation
Stride	~9.9–10 Hz sustained in soak harness
Encode pairing	Justified 60s chunks at ~30 fps with valid sidecar
Thermal	AI suspend threshold before encoder compromise
Storage	Prune-first when storage_health reaches CRITICAL

Soak, vision-verify, and unit test suites assert precomp contract stability. ONVIF PullMessages filters c=0 person events from the same detection stream.

Limitations and risks

• Device tracker: SNPE link failures on RVC4 may require host-side IoU tracking in the agent.
• Thermal: High die temperature suspends AI while keeping the encoder running—monitor MQTT die temp.
• Storage: Long runs without prune can mark recorder CRITICAL; justified metadata required for playback and cloud escalation.
• Privacy: Edge-only inference; cloud escalation uses signed object-storage uploads from the proxy only.
• Precomp immutability: Blur and quality gates in vision-verify complement but do not override on-device counts.

Related references

Topic	Doc
Full ML Pipeline hub	RF-DETR models, ML Pipeline & fleet deployment
Pipeline	Edge AI and RF-DETR
ONVIF person events	ONVIF reference
Cloud escalation	Cloud video escalation
NIST RMF mapping	NIST compliance
Proxy tools	API quick reference

Next steps

• RF-DETR ML Pipeline hub — training, DLC specs, exact fleet model names
• Justified recording — sidecar fields paired with detection stride
• Architecture diagrams — end-to-end dataflow
• Video understanding — cloud escalation after precomp

Operator depth

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