Pre-registration v2 (re-frozen 21 Nov 2026, Amendment 2): do GNN-proposed circuits plus polynomial post-processing need fewer parameters than a hardware-efficient ansatz? Test of a 4-qubit TFIM claim

Version 2 of the frozen protocol (v1: doi:10.5281/zenodo.24410738). The first scored run under v1 failed the pre-registered budget-equality gate G7 (scipy overshoot, max 10,030 evaluations vs a limit of 10,010) and is Inconclusive; Amendment 2, made after seeing those results, enforces a hard evaluation cap and changes nothing else; the scored run is repeated. Frozen protocol, code, labelled training set, trained GNN weights and pre-freeze gate results, deposited before any scored run. CLAIM (a machine-generated, untested hypothesis from solver.press): GNN-proposed circuit layouts with a 'channel polynomial' post-processing step reach diamond-norm error <= 0.01 with at least 15% fewer parameters than a hardware-efficient Ry-Rz-CNOT baseline, on a 4-qubit open-chain transverse-field Ising model at t = 0.5, 1, 2, while meeting every target the baseline meets. DESIGN: six arms (baseline, Trotter structure control, search-labeller control, GNN proposal, and polynomial post-processing on the baseline, the Trotter control and the GNN), 60 units per arm (3 targets x seeds 1000-1019), equal optimiser budget (10,000 gradient evaluations), exact diamond distance. DECISION RULE: Supported only if the combined method meets every target, its median parameter count is at most 0.85 of the baseline's, a one-sided paired Wilcoxon test gives p < 0.05, and the median shot-overhead factor is at most 10; an attribution line reports whether it also beats every control. SCOPE: only the row's own falsification clause is tested (the 30%, 12-qubit, MaxCut, SU(4) and hardware claims are not). The 'channel polynomial' is my operational definition, not the row's. PRE-FREEZE (gates G1-G6 re-run under Amendment 2): 500 labelled instances, five trained GNNs and gates G1-G6 (all pass) are included; no scored result exists.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-21
DOI
https://doi.org/10.5281/zenodo.24416326
Primary Topic
Quantum Computing Algorithms and Architecture
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Pre-registration v2 (re-frozen 21 Nov 2026, Amendment 2): do GNN-proposed circuits plus polynomial post-processing need fewer parameters than a hardware-efficient ansatz? Test of a 4-qubit TFIM claim

John Goodman
Zenodo (CERN European Organization for Nuclear Research)
Quantum Computing Algorithms and Architecture
article

Pre-registration v2 (re-frozen 21 Nov 2026, Amendment 2): do GNN-proposed circuits plus polynomial post-processing need fewer parameters than a hardware-efficient ansatz? Test of a 4-qubit TFIM claim

John Goodman
article en

Abstract

Version 2 of the frozen protocol (v1: doi:10.5281/zenodo.24410738). The first scored run under v1 failed the pre-registered budget-equality gate G7 (scipy overshoot, max 10,030 evaluations vs a limit of 10,010) and is Inconclusive; Amendment 2, made after seeing those results, enforces a hard evaluation cap and changes nothing else; the scored run is repeated. Frozen protocol, code, labelled training set, trained GNN weights and pre-freeze gate results, deposited before any scored run. CLAIM (a machine-generated, untested hypothesis from solver.press): GNN-proposed circuit layouts with a 'channel polynomial' post-processing step reach diamond-norm error <= 0.01 with at least 15% fewer parameters than a hardware-efficient Ry-Rz-CNOT baseline, on a 4-qubit open-chain transverse-field Ising model at t = 0.5, 1, 2, while meeting every target the baseline meets. DESIGN: six arms (baseline, Trotter structure control, search-labeller control, GNN proposal, and polynomial post-processing on the baseline, the Trotter control and the GNN), 60 units per arm (3 targets x seeds 1000-1019), equal optimiser budget (10,000 gradient evaluations), exact diamond distance. DECISION RULE: Supported only if the combined method meets every target, its median parameter count is at most 0.85 of the baseline's, a one-sided paired Wilcoxon test gives p < 0.05, and the median shot-overhead factor is at most 10; an attribution line reports whether it also beats every control. SCOPE: only the row's own falsification clause is tested (the 30%, 12-qubit, MaxCut, SU(4) and hardware claims are not). The 'channel polynomial' is my operational definition, not the row's. PRE-FREEZE (gates G1-G6 re-run under Amendment 2): 500 labelled instances, five trained GNNs and gates G1-G6 (all pass) are included; no scored result exists.

Zenodo (CERN European Organization for Nuclear Research)
Quantum Computing Algorithms and Architecture
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon