( 01 )Trust

Quality Gates & Measurement Methodology

Q: Are the thresholds the same for every scene family?

The gates are identical. Pass rates differ by family and are published per family on the data card rather than averaged into a single site-wide figure.

Q: What happens to clips that fail a gate?

They are rejected at ingest or flagged and excluded from the delivery. Nothing that fails a gate is silently mixed into a shipped batch, and the rejection rate is reported.

Q: Do you publish measured accuracy or only thresholds?

Both. Thresholds are published here; measured values for a specific delivery are on its data card, with a document number and a last-verified date.

Gates are enforced at ingest. Records that fail are rejected or flagged — never silently mixed in.

QUALITY GATES — ENFORCED AT INGEST

Hand visibility

≥ 50% of frames — Per-clip detector; below threshold the clip is rejected at ingest.

IMU–video drift

< 1 ms / 60 s — Hardware timestamps, verified per clip.

Reprojection error

< 3 px @ 1080p — Geometric consistency after 4D reconstruction.

Composite score

≥ 70 / 100 — Weighted across reconstruction, hand pose, tracking, retarget feasibility and simulation replay.

PUBLISHED ON EVERY DATA CARD

· Hours and rejection rate by scene family
· Hand-pose accuracy against the motion-capture subset
· Contact-label precision and recall on an adjudicated sample
· Rater agreement on L1–L3
· IMU drift and reprojection error distributions
· Collection regions and consent-withdrawal rate
· Retarget feasibility rate per end effector

Every card carries a document number, a revision and a last-verified date.

Data Rights & Provenance

Security

Licensing

Quality Methodology — Common Questions

Are the thresholds the same for every scene family?