Control of Error, Not a Gun to the Head: A Montessori-Informed Lesson Plan for Reinforcement-Learning Harness Design

This methods note proposes a Montessori-informed approach to reinforcement-learning (RL) evaluation-harness design, centered on Maria Montessori's principle of control of error. The paper examines the engineering consequences of incorrect answer keys, ambiguous evaluation items, and faulty reward signals. Its central proposal is that when a verifiably correct model answer is marked incorrect, human evaluators should audit and repair the answer key or checker before updating the model's policy. A six-session, 90-minute workshop is proposed for reward-function designers, preference-data labelers, evaluation-harness maintainers, and trace reviewers. The curriculum covers disagreement classification, answer-key auditing, separation of verifiable and ambiguous tasks, removal of erroneous training signals, reviewer conduct, and narrowly justified policy updates. A blank disagreement-log form is provided. The methodological framework draws on Montessori's The Montessori Method (1912), while separately examining her discussion of language and “supra-intelligence” in The Absorbent Mind (2012 eBook edition). Public statements concerning reinforcement-learning traces and the 2017 announcement of Saudi citizenship for the robot Sophia provide historical and contemporary context. These events are not treated as evidence of machine consciousness, suffering, personhood, or legal rights. Research status: This publication presents a proposed method only. No experiment has been conducted, no empirical result is claimed, and no institutional adoption is asserted. The manuscript underwent multi-architecture AI-assisted review on 12 October 2026, including review by Maelduin, Platin' Sleuth, Smitten, and Lyra. The author retains responsibility for the manuscript, source checks, interpretations, and publication.