The ledger

What we've measured, and on what.

Every figure here names its test set and the date it ran. These are lab tests on licensed public datasets and school-camera research video. The strongest numbers come first. The weaker ones sit lower on the page, each with plain context.

Engraved still life of an open ledger with a brass magnifier, calipers and a compass.
Plate I. Every figure with its test set and its date.
The headline figures

Where GroundsWatcher measures well.

0alarms
False alarms on school-camera video
2.835 camera-hours, 34 MEVA clips, after the two-looks rule. Nov 16, 2026.
12 of 12kept
Gun raises kept by the Eyes review
HBD21 held-out set. Nov 17, 2026.
0 of 15false yes
Scan false raises that Eyes passed on
From 34 MEVA school-camera clips. Nov 17, 2026.
14 of 15caught
Falls caught on CAUCAFall
With 0 of 15 false alarms on everyday clips. Nov 12, 2026.
The entries

Each test, in full.

The test set, the bar and every measure, in the order of strength.

Entry IRun Nov 17, 2026

Eyes review step, gun raises

Test set. HBD21 held-out set (38 clips, staged handgun video, CC BY 4.0), 140 other HBD clips, and the first scan's 15 false raises on 34 MEVA school-camera clips.

Review bar 0.5. Eyes looks at each raise from the first scan. It can lower an alert's priority, never delete it.

Eyes review step, gun raises
Gun raises kept, HBD21 held-out12 of 12
False yes on the first scan's MEVA false raises0 of 15
False yes on no-gun people, HBD21 held-out0 of 14
False yes on no-gun people, other HBD clips0 of 62
Entry IIRun Nov 16, 2026

False-alarm check on school cameras

Test set. 32,222 hand crops from 34 MEVA school-camera clips, 2.835 camera-hours (CC BY 4.0).

Every hand crop on these clips was scored. The two-looks-in-4-seconds rule turned the 10 single-crop yeses into 0 alarms.

False-alarm check on school cameras
Alarms after the two-looks rule0 in 2.835 camera-hours
Crops judged a gun at bar 0.70 of 32,222
Crops judged a gun at bar 0.510 of 32,222 (0.03%)
Entry IIIRun Nov 16, 2026

Phones and objects in hand

Test set. 2,290 full-frame crops of people holding phones and other objects, MEVA school-camera video.

Phones held in front of the body are the most common look-alike for gun detectors. These crops were scored the same way as live hand checks.

Phones and objects in hand
Judged a gun at bar 0.70 of 2,290
Judged a gun at bar 0.52 of 2,290 (0.09%)
Entry IVRun Nov 12, 2026

Person down, CAUCAFall

Test set. CAUCAFall test subjects (CC BY 4.0).

CAUCAFall uses the same room and camera as the clips the settings were tuned on, so it's the friendliest of these tests. It is not a school hallway.

Person down, CAUCAFall
Falls caught14 of 15
False alarms on everyday clips0 of 15
Where the numbers are weaker

The tests that didn't go our way.

They stay on this page. Each one says what was measured and why the figure is lower.

Entry VRun Nov 17, 2026

Eyes review, other HBD clips

Test set. 140 HBD clips outside the held-out set.

Eyes review, other HBD clips
Gun raises kept39 of 45

Below our own pass line of 42 of 45 for this step, so we count it as a miss. The held-out set above, which the settings never saw, kept 12 of 12.

Entry VIRun Nov 16, 2026

Gun check, scan plus full-detail hand look, without Eyes

Test set. HBD21 held-out set: 24 gun clips and 14 no-gun clips.

Gun check, scan plus full-detail hand look, without Eyes
Gun clips caught12 of 24
Alarms on no-gun clips0 of 14
First-generation detector alone, same clips4 of 24

This run measures the scan on its own, before the Eyes review step. It needs two looks at or above bar 0.5 on the same person within 4 seconds before an alarm. On this set that meant no false alarms and half the guns missed. It still caught three times as many as our first-generation detector on the same clips.

Entry VIIRun Nov 12, 2026

Person down, GMDCSA-24

Test set. GMDCSA-24 held-out subject (CC BY 4.0).

Person down, GMDCSA-24
Falls caught0 of 15
False alarms on non-fall clips1 of 22

A second camera with a close webcam view, where the fall rule tuned on CAUCAFall did not carry over. We publish it because a district should see where a rule breaks. Neither fall set is a school hallway.

Datasets: HBD21 handgun video (CC BY 4.0). MEVA school-camera video by Kitware Inc. and IARPA (CC BY 4.0). CAUCAFall, Universidad del Cauca (CC BY 4.0). GMDCSA-24 (CC BY 4.0). "Bar" is the score a check needs before it says yes.

An invitation

See it on your own cameras.

We bring the box, set it up on the cameras and doors you have, and walk your staff through it.

The careful eye on every campus.