Every claim below is a process we run, not a promise. This is the method behind the data you receive.
Every project starts with a reference set annotated by senior QC. The gold set fixes what a correct label looks like for your schema before any production work begins.
Analysts qualify against it, and it stays in circulation afterwards as the yardstick for drift.
A sample of every batch is independently double-annotated. Two analysts work the same items without seeing each other's output.
We report Cohen's Kappa per batch, so agreement is a number you can track across a project instead of a claim you have to take on trust.
We maintain a documented classification of domain-specific failure modes. Occlusion, identity switching, boundary disputes and similar categories each have a definition and examples.
Errors are filed against that taxonomy, and the resulting distribution drives training and review priorities.
Every delivery ships with per-batch metrics and error-trend notes. You see what was measured, what was corrected, and where the batch sits against previous ones.
The report is the deliverable, not a summary of it.
Every analyst signs an individual NDA. Client footage is access-controlled and segregated between projects.
Training runs through our internal academy with a certification exam. Analysts reach production work only after they pass it.
| Batch | Items | IAA (κ) | First-pass accept | Notes |
|---|---|---|---|---|
| B-014 | 4,820 | 0.93 | 97.1% | Baseline batch |
| B-015 | 5,140 | 0.91 | 95.8% | Occlusion-heavy period |
| B-016 | 4,960 | 0.94 | 97.6% | Post-review calibration |
| B-017 | 5,320 | 0.95 | 98.0% | Stable |
The numbers above are placeholders that show the shape of a delivery report. Your own batches report against your schema and your accept criteria.
Under NDA. No commitment.