demo recorder: auto-detect base model from adapter_config.json (1.5B-aware)

eval: FROZEN_LABEL env var to avoid overwriting existing baselines

demo recorder: PAIR_THEMES env var to also save as pair_N_theme_agent.json

eval: EVAL_TEMP env var to measure latent policy preference

eval: auto-detect LoRA base model from adapter_config.json

demo: add same-seed Frozen vs Trained recording launcher

support BASE_MODEL env var for 7B / larger model training

in-distribution eval support: EVAL_PHASE env var

phase-2 results: episode log + post-training automation

phase-2 fix: launcher was clobbering --reward_baseline when resume_from set

phase-2 tuning: anti-no-commit prompt + REINFORCE baseline env var

phase-2: SKIP_FROZEN + TRAINED_LABEL for re-eval against new checkpoint

phase-2: add resume-from-checkpoint to GRPO training loop

phase-g: T4 eval wrapper for Frozen + Trained Qwen on eval pool

phase-f fix: phase-1 domain coverage + checker robustness + flush + runs persistence

phase-f fix: drop pip install (UV venv has no pip; use PYTHONPATH=cwd)

phase-f: HF Jobs UV-script wrapper for T4 training

phase-d: 60 task templates x 6 domains + programmatic checkers + full loader

phase-b: paraphrase sampler + offline paraphrase generation script