Empirical validation

Constrained Reasoning Chains

An empirical telemetry study of LLM alignment under standardized ethical tracing

Eric Moore · 2026-04-28 · DOI: 10.5281/zenodo.19839280

Abstract

Turns consented reasoning traces into maps of completion corridors, hesitation zones, and refusal boundaries. Released alongside the open reasoning-traces dataset.

Key findings

• 6,465 traces analyzed for effective-dimensionality structure.
• An effective-dimensionality threshold around N_eff ≈ 7.1 correlates with successful conflict resolution; higher-dimensional reasoning runs show different qualitative behavior than lower-dimensional ones.