WORKING PAPER · VERSION 0.1 · 11 AUGUST 2026
Human Judgment, Institutional Error, and the Limits of Optimization
Zeyuan Li (Amy Li)
Founder & Independent Researcher, W-Axis Lab
Abstract
Artificial intelligence can improve an individual decision while weakening the resilience of the system in which that decision is embedded. When institutions rely on shared foundation models, overlapping data, common evaluations and similar interface defaults, errors may become correlated, minority signals may disappear, and apparently independent judgments may converge before uncertainty is resolved. The paper defines this condition as machine consensus, proposes measurable constructs and a four-condition pilot, and develops a W-Axis control stack around independent priors, protected dissent, provenance diversity, reversible action and accountable human authority.
Central proposition
An AI system can increase local decision quality while decreasing institutional resilience when adoption compresses the diversity, provenance, timing or reversibility of judgments.
Proposed measures
Dependency concentration, recommendation agreement, human judgment compression, counterfactual breadth, weak-signal retention, override integrity, detection lag, provenance diversity and correlated-error exposure.
Publication boundary
The full v0.1 manuscript is being held for author citation verification and substantive review. It is not represented as published or peer reviewed.
Research route
A request creates a private inbound record. It does not grant publication, quotation, reuse or affiliation rights.