finalization: blog + README + execution rewrite, drop 3B + openclaw shim

training: precision rewrite to prevent SFT collapse + GRPO variance starvation

hackathon sprint: grader collapse + coliseum rename + training pipeline

data: 30 expert episodes + training notebook

GRPO notebook: trajectory snapshots, Drive checkpointing, comparison table

Replace estimated baseline numbers with eval_sweep-verified ones

Auto-fetch seed dataset + Open-in-Colab badges

Wire Colab Secrets bridge into both training notebooks

Expand seed_combined.jsonl to 200 samples across 3 teachers

Add Groq driver, GRPO notebook, and eval-sweep harness

Add Llama-3.3-70B teacher episodes + merge into seed_combined.jsonl

Add Fireworks driver for teacher-data collection

Document the 4-step empty-prompt filter in seed README

Rewire sanity_run.ipynb to SFT on the 39-sample Claude seed

Add vibe-coded SaaS scenarios + Claude-teacher seed dataset

Harden env + ship Claude skill, OpenClaw-RL shim, training pipeline