i4h-workflow-train-rl: worth installing?

Reported results for i4h-workflow-train-rl (NVIDIA/skills) from NVIDIA, one row per receipt, in plain words.

Reported by: Reported by NVIDIA in its benchmark card
Model: GPT-5.5 (openai/openai/gpt-5.5, as the source names it)
Harness: codex, version not stated
Date: 2026-09-21
Tasks and runs: 9 tasks, as printed
Result: NVIDIA reported a score of 60.3% overall and no uplift figure (3 attempts per task). Overall: 60.3% — baseline ran, but no comparable score was available; uplift unavailable Security: 79.2% → 37.5% (-41.7 points) Correctness: 27.5% → 81.7% (+54.2 points) Discoverability: 77.1% — baseline ran, but no comparable score was available; uplift unavailable Effectiveness: 10.6% → 33.4% (+22.8 points) Efficiency: 71.6% — baseline ran, but no comparable score was available; uplift unavailable
Tokens: 15,261,479 tokens with the skill and 10,269,959 without, over all cases, as printed; the card gives no cost
Freshness: Reported on GPT-5.5. Freshness unknown: the registry has no release date for it.
Skill commit: 14a98ae278cec7e867695189d0f4003c08e4f765
Receipt: i4h-workflow-train-rl-openai-openai-gpt-5-5-imported-nvidia-cards-codex-2026-09-21-19b82bb498a4.json