WORKSPACE / METRICS

Performance metrics

PROJECT EVALUATION

Broader video evaluation

Results reported in the project review, dated 16 November 2026.

V1 / BROADER EVALUATION

Evaluated across a reported 142 unseen videos containing 100 snatching incidents and approximately 5.2 hours of normal footage. Incident detection and clip precision / recall use different denominators.

Incident detection
86%
86 / 100 incidents detected
Clip precision
87%
Reported clip-level precision
Clip recall
81%
Reported clip-level recall
Unseen videos
142
Approximately 5.2 hours of normal footage
Separate 23-incident threshold comparison

The following results come from a separate evaluation and are not part of the broader 100-incident result above.

Threshold comparison

Team-reported held-out test results · report Table 3
ConfigurationThresholdIncidents detectedFalse alerts / hourClip precisionClip recall
V1 Standard0.85419 / 23 · 83%691%78%
V1 High0.3020 / 23 · 87%1187%82%
V1 Maximum0.0521 / 23 · 91%1982%88%
Model B0.85411 / 23 · 48%1479%44%

High-setting clip counts: 71 true positives, 16 false negatives, 11 false positives and 522 true negatives. The report describes a compact normal test of approximately 0.29 hours.

Historical results and candidate coverage
Separate historical protocols · report Tables 4 and 5
EvaluationEvent recallNormal-footage alerts
Rules · Snatch 1.018 / 25 · 72%4 / 2.0 h · 2.0 / hour
Rules · Snatch-Web16 / 25 · 64%Not paired in the report
Fine-tuned verifier · Snatch-Web19 / 21 actor-matched · 90.5%3 / 0.286 h · 10.5 / hour

The historical verifier reported clip AUC of 0.96, an event-recall 95% interval of approximately 71–97%, and a false-alert interval of approximately 2–31 / hour. Different protocols and model-selection overlap prevent a controlled model comparison.

Candidate coverage increased from 74% to 92% across 50 labelled events (25 per dataset). Normal-footage candidate generation increased from 3,240 to 3,653 / hour. Coverage measures whether a candidate contains the labelled pair, rather than final detection accuracy.

Source: G024_Report.pdf · Section 5.1, pages 5–6. These are reported experiment results, not live measurements. The report notes that V1 source artefacts were unavailable for independent audit and live-camera performance was unverified.