Pre-registration (frozen 18 Nov 2026): does Newton-PINet reach its reported accuracy with 50 training tasks? Arm A of a replication of arXiv:2611.20609
Frozen protocol and runner, deposited before any production run. CLAIM (abstract of arXiv:2611.20609): on a nonlinear reaction-diffusion benchmark Newton-PINet achieves up to three orders of magnitude lower relative error than recent neural solvers with 16x fewer training tasks (Table 4: test relative MSE 1.72e-7 with 50 tasks, against 2.91e-4 for a baseline trained on 800). SCOPE: the baseline numbers are quoted from another paper's dataset and split, so the comparison itself is not tested here; Arm A tests whether the released code reproduces the accuracy with 50 training tasks, and how stable it is across seeds. DESIGN: the authors' public notebook (fycAIwj/Newton-PINet at commit d9987da6, notebook and data hashes fixed) is run unmodified apart from declared edits (CUDA line removed, seeds set), as shipped with seed 10 (twice, for determinism) and with ten fresh seeds, on CPU. DECISION RULE: Replicated if the as-shipped run and at least 8 of 10 fresh seeds reach test relative MSE <= 7.3e-6 (half the published log-gap to the quoted baseline); Did not replicate if the median of the ten fresh seeds exceeds it; otherwise Inconclusive. CONTROLS: data integrity, loss-function selection, determinism, a wrong-equation check (rho replaced by -rho must worsen the error 100-fold), and train/test split disjointness. A pre-freeze correction of a misreading of mine is recorded in the protocol. SMOKE: one 200-iteration CPU run under a throwaway seed established timing and installability; no metric was written or read. Arm B (re-running the baseline) is deferred and not part of this freeze.
Authors
- John Goodman (ORCID: https://orcid.org/0009-0008-4080-8043)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-18
- DOI
- https://doi.org/10.5281/zenodo.24340571
- Primary Topic
- Model Reduction and Neural Networks
- Type
- article
- Field-Weighted Citation Impact
- 0.00