Assurance
Quality
Quality here means what the validation pipeline has actually established, and nothing more. Checks that have not been built are named as such rather than quietly implied.
100%
Accepted for annotation
1 of 1 judged items passed every stage that ran.
0%
Flagged for human review
Usable but imperfect — clipping, long silences. Warned, never auto-dropped.
0%
Rejected
Unusable audio: silent, truncated, or below the minimum sample rate.
Validation stages
| Stage | State | What it establishes |
|---|---|---|
| Technical checks | Running | Duration, sample rate, clipping and silence, measured per asset. |
| Duplicate detection | Not built | Near-duplicate recordings and texts, so one submission cannot be paid for twice. |
| Language identification | Not built | Confirms the recording is in the language the task asked for. A mismatch warns; it never auto-rejects, because code-switching is expected and research-relevant. |
| Fraud signals | Not built | Replayed audio, implausible capture patterns, device-clock skew. |
The model measures; this platform judges. Thresholds live in configuration here, not in the model, so swapping the checker changes a version string and nothing about what counts as acceptable. Currently reporting: technical/1.0.0. When the model is unreachable, jobs wait — a model that is down never rejects a contributor's work.