[P2] Paper eval + figures

[duckyquang]

Phase: P2
Owner(s): Sparsh

---

Goal: Main paper results: generalization gap, perturbation, length extrapolation.

Where: marc/eval/paper/, paper/figures/, results/p2_main/.

Parallel track: Write run_main_eval.py skeleton Day 1 with placeholder checkpoint path. Swap in Stage-B weights when Quang finishes.

1. Implement run_main_eval.py + plot_results.py.

2. Run generalization gap, perturbation, length extrapolation suites.

3. Ablations: noise, guidance, purist reward — separate JSON per ablation.

4. Output paper/figures/fig_generalization.pdf, fig_perturbation.pdf, summary_table.md.

Done when: Summary table + 2 figures ready for paper draft.

Push to: PR sparsh/p2-paper-eval → main.

---

Auto-created from the MARC task board (index.html). Update the board if scope changes.