GRPO checkpoint from the TaskTrove X2 (upper PPO clip) hyperparameter sweep. Base model Qwen/Qwen3-Coder-30B-A3B-Instruct, trained on DCAgent/exp_rpt_multifile with SkyRL + Terminus-2. This is the clip-0.2 arm; the campaign verifier is pass_ratio shaping.

global_step_60 was selected by the trailing-5 reward EMA (alpha = 1/3) over the full restart chain -- the highest-EMA aligned checkpoint (EMA 0.226; step reward 0.229; pass@8 0.375; entropy 0.094) among saved exports (hf_save_interval 5; step 5 excluded as first save). Cross-validated by both the inline EMA and parse_skyrl_metrics.py --run_dir --save_every 5.

The clip-0.2 arm was terminated by the owner at step 78/80 (it was still advancing; the chain was scancelled). This is a healthy, still-improving run, not a collapse: reward trends up (~0.15 early -> ~0.20-0.27 by steps 46-78), pass@8 rises (~0.25 -> ~0.30-0.48), and policy entropy is stable to slightly declining (0.11 -> 0.083, 0.75x). A transient single-batch dip at steps 33-34 (reward -> 0) recovered immediately at step 35 and is not a collapse. Step 60 is the best saved checkpoint; the unsaved EMA peak was around steps 66-67 (the run kept improving past step 60).

See training_logs/ for metrics.csv, report.md, reward_plot.png, the resolved rl_config.json, and the gzipped .out chain.

Training-time Daytona/Harbor rollouts: open-athena/tt-x2_clip-hi0p2 (the last episode of each trial -- the rollouts the policy trained on after rollback/truncation).