repro-bundle/v1/scripts/06_update_memory.py

"""Append a record to MEMORY.md."""
addition = """

## Paper #15836 (bI9moH3UZw) - High-Dimensional Learning Dynamics of Quantized Models with STE - DONE, 3/3 CLAIMS VERIFIED
- OpenReview: bI9moH3UZw. arXiv: 2510.10693. Authors: Yuma Ichikawa (Fujitsu), Shuhei Kashiwamura (U. Tokyo), Ayaka Sakata (Ochanomizu & RIKEN AIP). ICML 2026.
- No official code. Re-implemented from-scratch in PyTorch + SciPy + NumPy from Theorem V.3 (ODE rhs), Proposition V.2 (macroscopic state functions m_psi/q_psi/r_psi), Proposition VI.1 (input-only closed-form fp), Theorem V.8 (small-eta prediction for joint weight-input quant), Eq. (9) (STE update).
- All 3 claims VERIFIED:
  - C1 (ODE convergence, Thm V.3): STE at d in {100,300,900,3000}, 10 seeds, tau_max=300. Asymptotic eps_g at d=3000 = 0.5005 vs ODE 0.5000 (rel err 1e-3). Late-traj RMS shrinks 0.019 -> 0.0072; log-log slope -0.29 (bound -0.5). Documented persistent early-transient offset of eps_g ~ -0.12 due to finite-tau non-Gaussianity of w_perp.
  - C2 (plateau + sharp drop, Figs 2+3): full-scale d=900, 5 seeds, tau_max=1200. Two-phase trajectory in all 9 (b,omega) settings. Plateau length 2x with omega. Drop sharpness max|d_eps/d_tau| grows 8x with omega. STE-ODE RMS <= 0.011.
  - C3 (fixed-point deviation, Thm V.8): closed-form matches STE (d=200, eta=5e-3, 2M steps) to <= 0.002 at every omega. Interior-p correction sigma2_psi*Delta^2*p*(1-p) observed exactly (omega=1.5: theory 0.5, STE 0.499).
- Compute: local RTX 5070 (12 GB), ~53 min total, $0 cloud.
- Unit tests: 7/7 PASS (quantizer moments vs hand-computed; STE vs unquantized linear-regression ODE max rel err 6.4%; macro functions degenerate correctly at s->0; input-only fp reduces to ridge limit; small-eta prediction monotone).
- GPU perf: replaced naive (x.unsqueeze(-1) >= theta).sum(-1) with levels[torch.bucketize(x, theta)] for the hard quantizer -- 6x speedup on RTX 5070. Made Figure 2 sweep tractable in 8 min instead of 80 min.
- Bucket: amkkk/quantized-ste-dynamics-repro-artifacts (synced to repro-bundle/v1: 36 files).
- Space slug: amkkk/quantized-ste-dynamics-repro. Publish SCHEDULED for 2026-07-21 03:08 via schtasks task repro_publish_bI9moH3UZw (58h delay). Did NOT call lb.publish() myself.
- Isolated workspace: repro_bI9moH3UZw/.trackio.
- trackio 0.31.5 needs huggingface_hub >= 1.24 (Volume import). pip install -U huggingface_hub to fix ImportError on first use.
"""

with open(r"X:\hf_challenge\MEMORY.md", "a", encoding="utf-8") as f:
    f.write(addition)
print("appended")