Document published UR-type action alignment concern and validation limits

README.md

@@ -47,11 +47,25 @@ This is a derivative release, with credit to the original RoboMIND authors.
 | UR5 | 42,005 | 4,736 |
 | Total | 121,101 | 13,436 |
 
-The subset contains successful demonstrations with usable video and state/action streams.
+The subset contains source-labeled successful demonstrations with structurally valid video and state/action streams.
 Identical recorded payloads are deduplicated; alternate task instructions are retained in
 `source_aliases` when present. Episodes with missing, incompatible, or frozen camera streams
 are excluded. This release does not contain every embodiment or episode in the upstream dataset.
 
+## UR5 action-quality caveat
+
+The [Qwen-RobotManip technical report, Section 2.4](https://arxiv.org/html/2606.17846#S2.SS4)
+reports that 81% of the RoboMIND UR-type episodes in its curation corpus failed a
+state-action trend-alignment check. The report covers both RoboMIND and RoboMIND 2.0;
+it does not identify which episodes in this derivative release are affected.
+**The reported 81% is not a measured failure rate for this release.**
+
+This release preserves the recorded leader-arm actions, without temporal correction.
+Its structural, deduplication, and frozen-camera checks do not establish causal
+state-action alignment. The 46,741 UR5 episodes have not passed a validated
+state-action alignment filter and should be excluded from action-supervised training
+until separately validated. Video-only use requires its own task-specific assessment.
+
 ## Format
 
 Files are grouped into uncompressed WebDataset TAR shards of approximately 4 GiB under