# research_log — WP28 "distributed MLE: local bias correction vs. global synchronization"

## 2026-10-28
- 18:33 Mandate received (brief `CLAUDE.md`, 1,055 lines). Series inspected; template and WP 25/27 conventions noted (PROJECT_MAP.md).
- 18:36 Prior draft located in `~/Downloads/WP28_UQO_Claude_Research_Brief/`; frozen under `references/initial_pamle/` with SHA256.
- 18:39 Original Monte Carlo re-run: `python3 monte_carlo.py` in a scratch copy, seed 20261028 → max |Δ| = 7e-15 (summary),
  4.5e-13 (replicates). Old `pamle.tex` compiles (16 pp.).
- 18:45 Cluster inventory: 22 nodes reachable; default `python3` lacks NumPy on most minis → Phase 8 will provision a venv per node.
- 18:47 Four literature-verification tasks launched (ZDW/Huang–Huo/Lin–Xi/Rosenblatt–Nadler; CSL/CEASE/FONE/DANE/EDSL/Battey/
  Volgushev/Shang–Cheng; jackknife & bias-correction classics; 2013–2026 priority-threat sweep).
- Decision: identifier WP28 kept (see SOURCE_PROVENANCE.md collision note).
- 18:55 Audit A received and archived (literature/audit_A_*.md, PRIOR_ART_MATRIX.csv rows 1-4). Key facts: HH19 proves one-step equivalence only under k=O(√N) from an uncorrected average; ZDW SAVGM = single-ratio subsampling correction with inflated variance constant, no jackknife/higher order; LX11 K≪N^{1/4}; RN16 m fixed.
- 19:00 Audit B archived (8 papers). Fixed sign of the analytic Cox–Snell correction (b_1 = +½ I^{-1} T; exponential check b_1 = θ0).
- 19:08 Audit C archived (jackknife classics; two DOI corrections; Dhaene–Jochmans partitions {2,3,..} bias-optimal; Hahn et al. 2024 variance of split-sample corrections). Patterns (1,2,3) added to the experimental plan.
- 19:12 Audit D archived (priority sweep; 23 more rows). Closest prior art for the combination: Zhu–Li–Wang 2021 (K=O(N^{3/4}) TWLSE), Gu–Chen 2023 (analytic debias+average, 2/3), Wu–Zhu–Wang 2023 (delete-one jackknife per block), Yin 2022 (bias-adjusted one-step, 2/3). No joint (r,q) frontier, no sharpness for corrected constructions, no quantized-Hessian statistical result found.
- 23:25 Laptop load: N=2^22 exponential cells relaunched with 500 reps (m=8 cell dropped: 14 h projected); exponential ablation remaining cells relaunched with 300 reps. Actual counts recorded in manifests/table notes.
- 23:25 Exponential ablation: 7 cells complete at 1000 reps, cell N=6291456 m=24 kept at its 300-rep checkpoint (not rerun).
- 23:33 Lean 4 / Mathlib formalization added at Simon's request: theory/lean (27 theorems, builds cleanly: Vandermonde, exact biases, Newton identities, basin, boundary constant, frontier exponent algebra, Gamma/Beta moments, exact Var J2).
- 05:29 Benchmarks: M4M64b's Python cannot open LAN connections (macOS local-network privacy: 'No route to host'); replaced by M3BA24 in all designs.
- 05:29 M3BA24 unreachable (asleep); the coordinator host M4M64a also serves as the 16th worker host in the benchmark designs (noted in the paper).