Immutable locator for the N1.7 ROBOCASA_GR1_TABLETOP benchmark checkpoint

At current main commit 376ba890cff8c9de64d71d982772a9c36185fdd7, examples/robocasa-gr1-tabletop-tasks/README.md reports 70% over 20 trials for:

gr1_unified/PnPBottleToCabinetClose_GR1ArmsAndWaistFourierHands_Env

using a finetuned N1.7 ROBOCASA_GR1_TABLETOP checkpoint. The evaluation command still contains <path-to-finetuned-gr1-tabletop-checkpoint>, while gr00t/data/embodiment_tags.py classifies ROBOCASA_GR1_TABLETOP as finetuning-only and not present in a shipped checkpoint.

Could you please provide either:

1. the official Hugging Face or NGC locator and an immutable revision for the checkpoint used for the published benchmark, together with its license; or
2. confirmation that the benchmark checkpoint is not distributed, plus the exact dataset revision, checkpoint step, and configuration used to produce the published 70% result?

The README currently provides the broad training command (NUM_GPUS=8 MAX_STEPS=60000 GLOBAL_BATCH_SIZE=512), but an immutable model artifact or exact final checkpoint identity would make the benchmark reproducible.