The previous previews were rendered per OLD segment and three of them (2026-09-14/episode_000_seg03, _seg04, _seg06) named segments that no longer exist -- the corruption-threshold fix merged those fragments into a single 15.2 min span. All 32 published segments now have one.
Remove the old toy segmentation superseded by the corruption-threshold fix
These 10 segments were cut when CAM_SPIKE_REL was 4.0, which read fast wrist-camera motion as a torn frame and split continuous recordings into fragments. The re-cut at 18.0 replaces them with longer segments covering the same source frames -- 2026-09-14/episode_000 went from 7 fragments totalling 4.9 min to a single 15.2 min span. Leaving both would put the same frame in two overlapping segments.
Toy-only incremental publication. Preserve existing parquets and splits; correct verified old index Z-up labels. Exclude recorder-invalid episodes and retain the existing 30-second clean-span policy. See data/toy/processing_20260919 for verification and provenance.
30 days ago
Synchronize V8 force for all218 episodes including validation; preserve accepted actions
Rendered with --force-root on the promoted v8 tree. For motherboard the
virtual-target overlay also reflects the repaired poses, since it reads the
release parquet; the projection circles still come from the raw H5, so a
frame where tracking glitched shows a jumping circle beside a smooth target.
That contrast is the repair being visible, not a calibration fault.
previews: 12 samples rendered against the v8 force channel
Three per task, rendered with --force-root pointing at the promoted v8
force tree. The remaining previews on the Hub still carry v5 overlays and
are being regenerated separately; these are for review of the new values.
