Oct 24 13 days ago 19:07 v1.3: eval.py --seed (the verifier draws a fresh one) 18:58 v1.3: task images pinned by digest (docker_image + Dockerfile FROM) 18:49 v1.3: task content hashes + snapshots of modified pool tasks; eval denominator (only pre-policy infra excluded, tail cut = fail); k up to 64; SGLang concurrency 64 14:24 v1.2: export refuses a file too large to ship as one transfer part 14:22 v1.2: verifier HF tier always private 14:08 v1.2: export writes hardlinked transfer parts (Modal 5 GiB file limit); W&B metrics-only; runner exit-hang guard 13:49 v1.2: runner kills leftovers after the trainer exits; stop keeps an ended run's state; pgid guard 13:45 v1.2: kill by run marker, eval stop/status, status after crash, checkpoint pruning + disk guard, live lr, W&B Oct 23 14 days ago 11:48 v1.1: task-container backend configurable (TASK_ENV_TYPE/APP/KWARGS), not coupled to Modal 09:00 code.tar.gz: eval k-per-task + early finish (T1-TRAIN) 08:25 code.tar.gz: task roots include /opt/task/data/tb21 + runaway guard (T1-TRAIN) 07:59 code.tar.gz: variants.py + eval.py failure path (T1-TRAIN) 07:42 code.tar.gz: max_seq_len config, GPU Adam TP4 layout (T1-TRAIN) 07:11 v1.0 layout: code / data / toolchain archives + MANIFEST under data_env_selection/; loose files removed 
