Shiki42/up-vla-precision-recovery-15k
Updated • 44
Protocol-registered E001 artifacts for endpoint, direct H16, and residual H16 recovery. The selected residual model improved grouped test path score by 4.68%, below the pre-registered 10% validation gate; E002 was not opened.
The source dataset includes RLBench-derived material with non-commercial research restrictions. Treat these checkpoints as research-only unless all upstream license obligations have been independently resolved. See experiments/E001/E001-CARD.md and the immutable run/evaluation receipts for details.