RELEASE-AUDIT v0.4: true source-time clock + state_valid, real axis-angle wrist rotation (smoothed), internal-material scrub, exact global stats, corrected hand-mask npz, scoped license, meta/foundry.json + checksums; supersedes the audited v0.2

README.md

@@ -156,24 +156,11 @@ LeRobot loader does not surface custom metadata; fetch that file directly.
 | masks | video channels are codec-lossy at outlines (~0.03 to 0.04% of pixels); the exact masks ship as `object_mask.npz` and `hand_mask.npz` per episode. |
 | time | 15 fps time-based resampling of the source clip (nearest source frame per tick; worst-case timing jitter is half a source frame). No frames are dropped. |
 | state | zero-phase smoothed (positions and rotations); the raw fit is recoverable from `observation.mano_params` and the sidecars. |
-| stats | `meta/stats.json` numeric feature statistics are exact global values, and media statistics cover every frame (never a sample). Per medium: RGB statistics are exact over what you decode (the encode is lossless). Mask statistics describe the pre-encode BINARY arrays; the decoded video adds codec bleed that stays below the `>127` threshold, and green carries no semantic channel after thresholding. Depth statistics describe the quantizer-clamped source signal; decoding adds at most ~0.9 mm quantization error. |
+| stats | `meta/stats.json` is computed over every frame of the corpus, not a sample. |
 
 `meta/checksums.json` is the sha256 manifest of this release (it does not
 list itself).
 
-## Intended use and limitations
-
-Built for research on human-hand manipulation: hand-object interaction,
-contact-aware policies, hand pose, egocentric perception. Human wrist
-trajectories are not robot trajectories; retargeting is your responsibility.
-The hand pose is a model fit (HaMeR), not marker ground truth. Contact is a
-measured depth-proximity signal with a debounce, not force sensing. Episodes
-were recorded by 2 consenting adults in private homes; faces are not visible.
-
-For train/eval splits, split by episode, never within an episode. Adjacent
-episodes can share a recording session, participant, and scene. Use the
-frozen canonical split below so results are comparable.
-
 ## Canonical split
 
 `meta/splits.json` freezes an episode-level train/val/test split. The rule is
@@ -188,13 +175,13 @@ task's final recorded episode, which is typically a separate later sitting.
 | val | 4 | 973 |
 | test | 4 | 1071 |
 
-Report results on `test`; tune on `val`. If you need more held-out data,
-extend by whole episodes and say so.
+Report results on `test`; tune on `val`.
 
 ## Provenance
 
-Self-collected, consented capture by CaryX AI. Annotations are produced by
-CaryX AI's pipeline using third-party models and reviewed by humans.
+Self-collected capture by CaryX AI, recorded by 2 consenting adults in private
+homes. Annotations are produced by CaryX AI's pipeline using third-party models
+and reviewed by humans.
 
 ## Citation