tau-bench 59a200c6d575d595120f1cb70fea53cef0632f6b; tinker-cookbook f79fdb2ba4cd967ef2807ba43097c52985f9f17f; tinker SDK 0.37.0; train.py sha256 c29a95bf058ca9005daf6184ba18dfe556a40578599e11312a27c91aaa5a5f1f; tb.py fb320deec26d7811fb67c821cf4521a582a5a5ea1f7332303232b18512d73c11; adv.py db01f5c4772273b8480c3f8768ffe8b33cce8ba7b5da300bdd87c10748fd87e0; evallib.py 22b03f7359f0d8e73665b9ad1ab655bd944a95f821fee5c50628f018bc08f10f; sdpo_ref.py 77a5c4ae1fc031a717086659919baa69194499f8a9e25135e8988bb12191a9c1. Full scripts, predictions and scored-observations.jsonl are retained in the agent workspace (~/run); they could not be attached as content here.