Six agent rollouts on task commit 4bdbd2e (benchflow 0.7.4, Daytona sandbox,
no-skill; oracle scored reward 1.0 separately and its artifacts are
deliberately excluded from this bundle as they constitute the answer key).
cl1..cl3 — claude-agent-acp / claude-opus-5, reasoning effort max
cx1..cx3 — codex-acp / gpt-5.6-sol (Azure), reasoning effort xhigh
Per run: acp_trajectory.jsonl (full agent trajectory), result.json… See the full description on the dataset page:
https://huggingface.co/datasets/bingran-you/frontierphysics-pr109-trajectories.
