Plain words in.
Verified alerts out.
Tell it what worries you. It works out what to watch, proves the rule, and has a second model check every alert.
The demo replays a recorded run with no model calls (labelled REPLAY), so it works with every GPU switched off. The verify page recomputes every SHA-256 in your browser.
“Keep people at least 4 metres away from a moving forklift.”
2.90m
Measured on the calibrated floor. The rule says under 4 m.
A person is shown within about 2.9 m of an orange forklift that is moving, so the 4 m separation requirement is not met.
- Muse Spark 1.3planned the rule133.6 s
- SAM 3.1measured the gap
- Muse Glimmer 30Bchecked the frames12.9 s
The system right now
From a worry to a proven rule
Nothing runs until a person approves a plan that passed every check.
- Step 1, not yet:
Describe
Say or type what worries you, in your own words and language. A camera's limits are said up front.
- Muse Voice Transcribe
- Muse Spark 1.3
- Step 2, not yet:
Plan
Muse Spark plans the job. SAM 3.1 looks at the camera, measures distances and checks it can see what the rule needs.
- Muse Spark 1.3
- SAM 3.1
- Step 3, not yet:
Prove
Rule CI attacks the rule with edge cases and broken copies before anyone can approve it.
- Rule CI
- Muse Imageoff
- Step 4, not yet:
Trust
A second model checks every alert on the video frames. Every decision goes into a hash-chained log you can check yourself.
- Tier 1 check
- Tier 2 review
One engine, three worlds
The same planner, Rule CI and verifier, pointed at three kinds of site.
Stock footage, Pexels 39221979 
Stock footage, Pexels 32243654 
Research footage, Jülich
Which Meta model does what
What runs in the deployed app today, said plainly.
| Model | Job | Today |
|---|---|---|
| Muse Spark 1.3 | Plans the job, compiles the rule, answers the owner's questions (Ask) | Live |
| Tier 1 check | Checks every alert on the video frames before anyone is bothered | Checking the API… |
| Tier 2 review | Second opinion: worker disputes, the plan audit and Ask's answers | Checking the API… |
| SAM 3.1 | Finds and tracks people and objects; probes what each camera can see | Checking the API… |
| Muse Voice Transcribe | Spoken requests | Live |
| Llama Prompt Guard 2 | Screens text seen in the scene before it reaches the checker | Wired; degraded in this deployment |
| Muse Image | Test scenes for Rule CI | Built; switched off in the app |
| Muse Code | Writes a new check when the rule language cannot say it; runs only after the owner approves | Built; owner-only (Extensions) |
The recorded runs on this page (1 and 2 Nov 2026) were checked by Muse Glimmer 30B (open weights) on our own GPUs.
Measured, not promised
Each figure is read from a committed result file. Weak results are published too: the README results table.
2.90 m
A worker passed a moving forklift; the rule said 4 m. Glimmer checked the frames: violation, 0.99.
live run, 1 Nov, T02reports/eval/2026-11-01-live-test-campaign.json
0.05
The rule engine flagged the seated driver. Glimmer saw it was the driver, so no alert went out.
live run, 1 Nov, T04reports/eval/2026-11-01-live-test-campaign.json
19.0 s
From the event to a checked decision, on the deployed stack.
live run, 1 Nov, T02reports/eval/2026-11-01-live-test-campaign.json
$0.0039
GPU cost per alert check, 20 checks at once on warm serverless GPUs.
load test, 2 Novreports/eval/2026-11-02-serverless-load.json