Topological Ethics paper 2

This technical report is the companion to the concept paper "Topological Ethics: A Geometric Framing for Alignment Regularizers" and documents its four gridworld experiments in full. Each experiment tests a penalty that protects a simulated human's option-space, the cells it can reach. The report gives the methods, every setting with its exact optimum, and the hypotheses with when each was stated. It also includes a disclosed design history covering an invalid early experiment, a discarded layout, and two bugs that were found and fixed. In brief: the penalty kept agents from trapping a stationary human; a candidate formalization of "torsion" added nothing; penalties checked only at periodic audits missed restrictions between checks; with a goal-seeking human, the penalty achieved safety mainly by avoidance; and cooperation appeared only when the penalty encoded a forecast of the human's next one or two moves. Code and data, which reproduce every reported number, are at doi:10.5281/zenodo.23841251. AI disclosure: The experiment code, analyses, and editing were carried out with the assistance of Claude (Anthropic).