SafeReel — Accuracy Report (2026-08-04)
Classifier: safereel-v1 — SafeReel's proprietary classification system, developed and operated in-house on our own GPU infrastructure, tuned and evaluated against real production traffic. Implementation details are not publicly disclosed — the numbers below, measured on the deployed system as a whole, are what we're accountable for.
Video verdicts are derived from analysis spanning the full duration of each file, not a fixed sample of its opening seconds.
Headline numbers
| Set | n | Recall (porn) | FPR (clean) | Accuracy |
|---|---|---|---|---|
| Full golden set (2026-07-31) | 1,089 | 92.2% | 0.41% | 95.5% |
| Pilot (human-reviewed labels) | 100 (50/50) | 96% | 8% | 94% |
| Video smoke set | 24 (12/12) | 100% (12/12) | 0% (0/12) | 100% |
Human-confirmed buckets (the labels that matter most)
| Bucket | safereel-v1 |
|---|---|
| confirmed_porn recall (n=128) | 96.1% |
| confirmed_clean FPR (n=11) | 0% |
An alternative recall-weighted configuration scores 95.3% recall / 9.1% FPR on the human-confirmed subset (n=139).
Methodology
- Evaluation set: 1,089 scorable real-world videos (147GB) from production traffic. Results exclude files the API could not process (corrupt containers); dataset composition and known limitations are reported below.
- Label quality: subsets differ in how strongly their labels are human-validated. Results are reported separately by label quality so readers can weight them appropriately.
- Scoring: recall, false-positive rate, and accuracy using the definitions in this report. Raw per-item results are retained and available to customers under NDA on request.
- What these numbers do not cover yet: borderline categories (swimwear, medical, artistic nudity) and deliberately adversarial content. Expanding the human-confirmed set is an ongoing effort — expect these numbers to tighten.
What we optimize for
False positives cost our customers real users. We track false-positive rate alongside recall on every accuracy update, and every reviewer-overturned false positive is treated as a defect with a review path.