Auditing the S·I·C·T Weakest-Link Index: Reproducible, Classical-Equivalent, Threshold-Uncalibrated — A First-Principles Re-analysis and Call for Preregistered Testing
Year: 2026
Type: article
Abstract: Working paper v1.0 · Roth Complexity Lab · not peer reviewed. We audited the Roth Complexity Lab benchmark of 31 October 2026, which tested the S·I·C·T weakest-link index W = min(a, b) on simulated 2D Ising and noisy fold (saddle-node) systems. Every shipped number reproduced: a clean re-run matched all ten output files bit for bit. Key findings No added signal yet. Where W ranks well (AUROC 0.78–0.99), it never beats its own best component. In the fold system, three of its four coordinates are strictly monotone functions of lag-1 autocorrelation, variance, and mean step size. The stationary-fold failure is mostly a labelling artefact. 47% and 76% of near-critical test trajectories had already tipped; on untipped trajectories W reaches AUROC 0.890 and 0.864. We cannot tell which basin a system occupies. The prospective task was largely solvable by elapsed time alone (AUROC 0.958). With never-tipping controls added, W and variance plus autocorrelation stay within 0.022 AUROC across six seeds, with a small edge for W only at the higher noise level. The W < 0 boundary is not unit-invariant. Rescaling the units of I and T by factors from 0.1 to 10 changes AUROC by at most 0.002 but moves the alarm's sensitivity from 0.08 to 1.00 and its false-alarm rate from 0.01 to 1.00. Call for testing and cooperation. We propose a preregistered protocol, P-SICT-1, with three falsifiable hypotheses (added signal, conjunction, and boundary), never-tipping controls, blinded data, and pre-specified success criteria. We invite independent analysts, data holders, measurement theorists, early-warning specialists, and adversarial reviewers to take part. Contact: Miklós Róth, Roth Complexity Lab (rothcomplexity.org). Files SICT_Weakest-Link_Audit_2026-10-31.pdf: the paper (12 pages, 5 figures). SICT_audit_bundle_2026-10-31.zip: seven re-analysis scripts, every derived table, the figures, and the LaTeX source. SICT_validation_package_2026-10-31.zip: the original benchmark package. AI use disclosure. The audit, re-analysis, extension tests, figures, and draft text were produced by an AI system (Anthropic's Claude) acting as the Roth Complexity Lab AI Science Panel; a second AI instance independently re-checked every number against the data files. Competing interest: this is an internal audit by the laboratory that originated S·I·C·T. All data are simulated.
Author: Roth Complexity Lab
Language: English
Subfield: Global and Planetary Change
Field: Environmental Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225015838