Documentation / THOR Data
14 / 22
Verification layers
Different checks catch different kinds of bad work.
Automated gates can reject missing coverage, gaps and malformed timing. Hidden repeats expose low-effort submissions, agreement between suppliers reveals outliers, held-out models test whether labels improve training, and stake-backed trust can penalise repeated abuse.
| Number | Layer | What it catches | Status |
|---|---|---|---|
| 01 | Automated task gates | Coverage, gaps, duration and timing errors | Spec defined |
| 02 | Hidden gold / repeats | Lazy labels and low-effort submissions | Design target |
| 03 | Multi-supplier consensus | Outliers, disagreement and coordinated noise | Design target |
| 04 | Held-out model evaluation | Labels that look plausible but do not improve training | Research |
| 05 | Trust + validator stake | Repeat abuse, collusion and rubber-stamping | Design target |
Automated gates are defined for the first task. The other layers are design targets or still in research, and each adds a check a low-quality submission has to pass before it is paid for.