From Metaphor to Loop: A Human-in-the-Loop Protocol for Autonomous Self-Improvement in AI Anagram Generation (Educational Transformation 003)

Overview By assigning a large language model three internally separated roles- a generator, a judge, and a prompt engineer- and connecting them in an iterative loop, strict anagram generation improved autonomously across cycles, while the human user retained control over when and how the loop advanced. This report is a sequel to Educational Transformation 001, "When Invoking Quantum Mechanics and Crystal Growth Metaphors, AI's Anagram Generation Ability Leaped Forward" (DOI: 10.5281/zenodo.18187938). Where the previous work asked how physical metaphors could draw out an LLM's latent precision within a single response, the present work asks what happens when the model is allowed to revise its own operating protocol, under human supervision. I am not an expert in AI or linguistics. I am a faculty member specialising in molecular biology (RNA) in the Educational Transformation (EX) division at the University of Tokyo, where I develop liberal arts curriculum and materials for first- and second-year undergraduates. This work, like its predecessor, emerged from teaching practice rather than from AI research. The "Quantum Anagram Self-Improvement Protocol v3.1" presented here structures a single LLM session as a multi-agent system: an Artisan generates anagrams under the current protocol; a Judge scores the best candidate on four fixed axes (constraint fidelity, semantic quality, aesthetic value, and surprise) and must always identify a concrete weakness; a Prompt Engineer then revises the generation protocol itself, never the scoring criteria, in explicit diff format. Each cycle's best anagram is logged into a cumulative ranking table, so that the trajectory of improvement remains visible at a glance. Crucially, the loop is not fully autonomous. The human decides whether to run one cycle at a time or several in succession, when to stop, and when to accept convergence (declared when the overall score exceeds 95 points); a rollback rule forces the Engineer to change course when scores decline for two consecutive cycles. The design is thus an exercise in human-in-the-loop orchestration: the AI improves itself, but the tempo, the taste, and the termination remain human decisions. This article is a practical report on prompt engineering as a liberal arts material, an attempt to let students observe, steer, and critique an AI's self-improvement process, and, in doing so, reflect on where human judgement should sit within autonomous loops. 概要 あ Japanese Abstract 私はAIや言語学の専門家ではありません。東京大学で「教養教育(リベラルアーツ)」の教材開発に携わる分子生物学(RNA)を専門とする教員で、Educational Transformation(EX)部門に所属しています。本稿は前報 Educational Transformation 001「AIに物理学的(量子力学と結晶成長論)メタファーを憑依させたら、アナグラム生成が劇的に改善した話」(DOI: 10.5281/zenodo.18187938)の続報・改良版です。 前報が「物理学的メタファーによって単一応答内のAIの論理性能を引き出す」試みであったのに対し、本稿では「AIに自らの動作プロトコルを改訂させるループ」を人間の監督下で回すことを試みました。 本稿で提示する「量子アナグラム自己改善プロトコル v3.1」は、1つのLLMセッション内に3体のエージェント——生成担当の「職人」、固定された4軸(制約適合・意味性・美的価値・意外性)で厳格に採点し必ず弱点を指摘する「審査員」、その指摘に基づき生成プロトコル自体を差分形式で改訂する「プロンプトエンジニア」——を分離して配置し、生成→審査→改訂のサイクルを反復させるものです。各サイクルの最良アナグラムは累積ランキング表に記録され、改善の軌跡が一望できます。 重要なのは、このループが完全自律ではない点です。サイクルを1回ずつ進めるか一気に回すか、いつ止めるか、収束(総合スコア95点超)を受け入れるかは人間が決定し、スコアが2サイクル連続で低下した場合のロールバック規則も設けています。すなわち本設計は、AIが自己改善する一方で、テンポ・美意識・終了の判断は人間の手に残すという Human-in-the-Loop の実践です。 本稿は、学生がAIの自己改善プロセスを観察し、操縦し、批評すること、そして自律的なループの中で人間がどのように介入するべきなのかを検討する方法の確立、を目的とした、教養教育教材としてのプロンプトエンジニアリング実践経過報告です。 Acknowledgements I am deeply grateful to Professor Tom Gally (Project Professor, Center for Global Education, The University of Tokyo; https://www.gally.net/), whose regular face-to-face sessions of playful experimentation with AI have given me numerous insights. The spirit of play that runs through those sessions — treating generative AI not merely as a tool but as a space for intellectual curiosity — underlies the design of this work. (謝辞)東京大学グローバル教育センター特任教授の Tom Gally 先生に深く感謝いたします。定期的な対面セッションでの遊び心(playful)に満ちたAIとの"遊び"は、私にさまざまな気づきを与えてくれています。生成AIを道具としてではなく知的好奇心の場として扱うその精神が、本稿の設計の根底にあります。 Disclosure of AI Use This document was prepared and edited with the assistance of artificial intelligence tools, including Google Gemini, ChatGPT, and Claude. The "Quantum Anagram Self-Improvement Protocol" described in this work — a generate–judge–revise loop with human-in-the-loop control — was itself designed, tested, and revised in dialogue with these tools, and thus constitutes both the subject and the product of the experiment. It represents a continued attempt, following Educational Transformation 001, to harmonise Western reductionism with aesthetic sensibility, extending it here to the question of autonomy and human judgement.