README: add Training History section — 3,200 episodes across 6 GRPO runs

Notebook: add Phase 8b — training history (3,200 episodes, 6-run ablation table)

Notebook: drop bitsandbytes (triton 3.x ABI break) — fp16 inference doesnt need it

Notebook: fix PYTHONPATH inheritance for !python -m chaosops subprocesses

README: signpost the phase-wise judge demo notebook

Phase-wise judge demo notebook (rubric-aligned, 25 cells, Phase 3A LoRA)

Ship Colab training notebook so README link resolves for validation pass

Promote Phase 3A LoRA — Qwen 3B beats heuristic on HARD, 100% rogue catch

Phase 3A: learning_curve.png

Phase 3A: comparison_curve.png

GRPO: add --rogue-bonus-multiplier to amplify oversight gradient signal

Promote Phase 2 LoRA (curriculum + LR=2e-5 + r=32) as the live trained lane

Phase 2 LoRA: learning_curve.png

Phase 2 LoRA: comparison_curve.png

Add A/B comparison Job for trained-policy showdown

GRPO: expose --learning-rate, --temperature, --curriculum-schedule

README: real before/after numbers from the 540-episode evaluation

Ship trained-policy artifact: training_metrics.json

Ship trained-policy artifact: learning_curve.png

Ship trained-policy artifact: evaluation_summary.txt

Ship trained-policy artifact: comparison_curve.png

Add HF Jobs entry-point for post-training evaluation

Ship baseline_curve.png so the Space README embed renders

README: add submission links, composable-rubric docs, plot embeds, package layout refresh

Add transformers-backend GRPO loader (no triton/Unsloth dep) + fix Jobs deps

Add HF Jobs entry-point script for GRPO training

Phase A submission cleanup — OpenEnv compliance + composable rubrics + loud-fail trained lane

rebuilding

readme

Initializing space