## Trigger

- Planning a research reproduction or training run with many GPU jobs, seeds,
  tasks, methods, categories, or ablation arms.
- About to write a new shell/Python loop for GPU dispatch, batch running,
  skip-completed behavior, per-job logging, or device-pinned experiment runs.
- The task mentions SafeAct-Steer, OpenVLA, LIBERO, remote GPU runs, or
  `CUDA_VISIBLE_DEVICES`.

## Fix

- First inspect the SafeAct-Steer reference implementation, then adapt the
  design into the target project:
  - GitHub repo: `https://github.com/lihan3238/SafeAct-Steer`
  - Branch observed when captured: `paper/anchors-and-channels`
  - Reference runner scripts:
    - `experiments/run_f2.sh`
    - `experiments/run_f3.sh`
    - `experiments/run_ablation.sh`
  - Reference experiment device recording:
    - `experiments/15_openvla_goal_adaptive_precommit.py`
- Reuse the pattern, not the private runtime state:
  - explicit GPU argument or `CUDA_VISIBLE_DEVICES`;
  - `set -euo pipefail` for shell runners;
  - one log per run under a project-owned log directory;
  - explicit report/rollout names;
  - source-device recording in output metadata;
  - project-specific checkpoint, env, candidate, seed, and output naming.
- Before launching work, still follow `lihan-cards` research mode:
  - read only the needed research profile sections;
  - use authorized GPU IDs from the profile, not visible `nvidia-smi` alone;
  - do not copy raw logs, result files, checkpoints, private paths, or remote
    machine state into public cards.
- Do not create or install a new global skill unless this becomes a repeated
  cross-project workflow with a stable reusable template.

## Reuse Rule

- Load before creating or reviewing GPU batch runners for research
  reproduction/training.
- Do not load for one-off single-command GPU runs, simple paper-reading tasks,
  or experiments already managed by a mature scheduler such as Slurm/Ray.
- If SafeAct-Steer paths have changed, verify the current GitHub repo first and
  update this card before applying stale paths.