Measured evidence · Jetson AGX Thor

Physical AI Safety Observability

Every number on this page is read from a run artifact under reports/thor/; each row links to its JSON. Device: jetsonthor, tegra264, # R38 (release), REVISION: 4.0, GCID: 43443517, BOARD: generic, EABI: aarch64, DATE: Wed Dec 31 00:15:19 UTC 2025, NV Power Mode: 120W 1. Runs on 2026-09-09, repository commit 254ef27. These are runtime-overhead and inference-cost measurements on simulation footage; none is a detection-quality number.

28.588 frames/sworker at 30 fps pacing, mock model, rule evaluation p95 0.99 ms

2.69 s p50Cosmos-Reason2-2B per frame under a JSON-schema grammar, 0 out-of-vocabulary labels, 6 phantom persons

3.2 s p50Cosmos-Reason2-8B under the same grammar, 0 phantom persons, 0 false events

not measuredprecision and recall: no labelled ground truth exists yet

Runtime overhead, mock adapter

The mock adapter returns fixed detections at microsecond cost, so this is the price of frame sampling, the policy engine, event construction and the backend path. Events are four per frame by construction.

Run	Frames	Frames/s	Rule eval p95, ms	Backend POST p50, ms	VIN p50, mW	Peak RSS, GB	Events
mock, 30 fps pacing, no posting mock_30fps_no_post.json	1,800	28.588	0.9877	n/a	24,604	0.061	7,200
mock, 30 fps pacing, per-event POST to FastAPI + SQLite mock_30fps_backend.json	400	4.357	0.7123	160.9	25,570	0.051	1,600
mock, 30 fps pacing, one batch POST per frame mock_30fps_backend_batched.json	400	7.075	0.8565	68.2	25,158	0.057	1,600
mock, unpaced, no posting mock_no_post.json	500	3,077.556	0.2257	n/a	n/a	0.041	2,000

Real-model inference cost, Cosmos-Reason2 via vLLM on the same device

Same 60 frames (one every 60 from an Isaac Sim recording with no people in it), same server image, same device. Inference in seconds per frame. "Phantom persons" are detections labelled person on people-free footage; "events" are the safety events those detections fired, all false by construction.

Run	Frames	p50, s	p95, s	max, s	Frames/s	Completion tokens p50	VIN p50, mW	VIN peak, mW	Detections	Out-of-vocab labels	Phantom persons	Events
Cosmos-Reason2-2B, think block, 4096 cap cosmos2b_video_60.json	60	4.35	12.85	15.41	0.154	260	66,568	107,102	97	59	0	0
Cosmos-Reason2-2B, no think, 512 cap cosmos2b_video_60_nothink.json	60	3.57	6.58	9.35	0.241	157	60,338	105,206	106	57	0	0
Cosmos-Reason2-2B, grammar, original prompt cosmos2b_video_60_schema_v1prompt.json	60	0.31	0.33	4.19	2.303	7	37,052	101,448	0	0	0	0
Cosmos-Reason2-2B, grammar, JSON-only prompt cosmos2b_video_60_schema.json	60	2.69	3.36	8.03	0.367	135	62,522	83,120	193	0	6	7
Cosmos-Reason2-8B, no think, 512 cap cosmos8b_video_60_nothink.json	60	8.29	27.78	33.67	0.098	126	69,136	108,376	81	20	0	0
Cosmos-Reason2-8B, grammar, JSON-only prompt cosmos8b_video_60_schema.json	60	3.2	8.26	14.96	0.234	47	65,976	131,582	74	0	0	0

Per-frame inference latency, p50 and p95, seconds. Reasoning length drives the tail; the grammar removes most of it.

Counts over 60 frames. Zero out-of-vocabulary labels under the grammar; the 2B then reports people that are not there, the 8B does not.

Achieved frames per second against a 30 fps source; the synchronous POST to the backend is the bottleneck, batching lifts it from 4.4 to 7.1.

Board input power at p50 over each run, watts; about 24 W is the device's idle level.

What is not established

No precision or recall for any rule: there is no labelled ground truth. The grammar fixes the vocabulary; only ground truth can score the detections.
No real camera input in these runs; the footage is simulation.
Runs are 60 to 390 seconds, not sustained operation.
Sixty frames of one video is a signal about the 8B, not a rate.

Generated by scripts/build_evidence_pages.py from the committed artifacts. Repository: https://github.com/obiedeh/physical-ai-safety-observability. Case study: obiedeh.github.io.