README.md

metadata

title: 'Finite-width NTK: claim-by-claim reproduction'
emoji: 🎯
colorFrom: yellow
colorTo: red
sdk: static
pinned: false
tags:
  - trackio
  - trackio-logbook
  - open-experiment
  - icml2026-repro
  - paper-SOlPHMdSY3

Finite-width NTK: claim-by-claim reproduction

Current verification supersedes the Historical rejected baseline. Start with the canonical evidence index. Every claim below is backed by executable source, inline numbers, raw JSON, an independent checker, and a negative control.

Paper: Finite-Width Neural Tangent Kernels from Feynman Diagrams, v4. Source archive retrieved 2026-08-01 with an explicit User-Agent; SHA-256 5237b9fec2f128b23266771acbc8e44837619d5309cd32840ce537071b64c47f.

Claim	Verdict	Canonical evidence
1. Section 4 graphical rules determine the order-1/n layer recursions	VERIFIED	Claim 1
2. The first NTK-mean correction contains exactly five diagrams	VERIFIED	Claim 2
3. Scale-invariant ReLU/LeakyReLU diagonal means have no finite-width correction	VERIFIED	Claim 3
4. Source Figure 2 GeLU data agree with the first-order correction	VERIFIED	Claim 4
5. Critical C_W=2 NTKs scale linearly through depth 30	VERIFIED	Claim 5

Fixed command:

uv sync --frozen --all-extras && uv run --frozen python -m reproduction.run_all

The cumulative evidence run used Hugging Face cpu-upgrade: estimated 8 cores; actual cgroup quota 8 CPUs, 32 GB flavor contract, no GPU; scientific runtime 5,225.662 s. Its exact raw output is downloadable JSON. The independent release rerun at Git SHA 0955d05f3d38f60adc356c3ea8fcf4fce3adb2b2 used the same allocation and command, took 5,363.686 s of scientific runtime, and matched Claims 3–5 exactly (max combined-SE z = 0.0, threshold 5.0). Download its terminal summary.

Method, environment, and limitations · Release report and visibility matrix · Historical rejected baseline