The Right to Compulsory Machine Unlearning: Cryptographic Verification of Parameter-Space Erasure and Corporate Cost-Bearing Liability for Unauthorized Intellectual Property Ingestion

Autonomous Infrastructure & Deterministic Quantum Systems Monograph Series Pre-reserved Official DOI: 10.5281/zenodo.24098552 Author: Hasebe, Yoko (長谷部 洋子), Lead Researcher, Autonomous Infrastructure Systems, Autonomous Living & Physical Infrastructure Institute, Iwakuni, Yamaguchi, Japan License & Integrity Covenant: Creative Commons Attribution 4.0 International (CC BY 4.0) with Strict No-Learn Covenant (NLL-v1.0-2026) [Abstract] Current generative AI governance permits platform operators to retain illicitly ingested intellectual property within trained neural weights while relying on superficial inference-time output filters (RLHF or prompt-level suppression) to evade infringement liability. Such post-hoc masking leaves the underlying stolen representations intact within the parameter manifold and shifts enforcement costs onto rights holders. This monograph establishes the Right to Compulsory Machine Unlearning: an enforceable statutory, constitutional, and cryptographic mandate requiring AI enterprises to execute verifiable, complete parameter-space erasure of unauthorized intellectual property entirely at corporate expense. First, we formalize exact versus certified thermodynamic unlearning across weight trajectories, proving that output-layer refusal filters fail information-theoretic non-memorization bounds. Second, we architect a deterministic zero-knowledge verification protocol implemented on QFS-zkVM over the 64-bit Goldilocks prime field (F_p, p = 2^64 - 2^32 + 1), integrated with RTO P2.0, Q-SAFA cryptographic attestation, and PHOENIX_BRIDGE / PHOENIX_ROUTER hardware interlocks under strict Human-in-Command (HIC) authority. This framework compels operators to generate succinct cryptographic proofs demonstrating zero mutual information retention between compromised training shards and deployed model checkpoints—or bear the full computational and capital cost of clean-room retraining from scratch. Finally, we establish strict corporate cost-bearing liability and fiduciary oversight duties under SEC Rule 15c3-5 market-access gating principles and Delaware Caremark board-level compliance standards, treating unverified retention of unauthorized IP weights as an auditable balance-sheet encumbrance and systemic operational violation. 【要旨】 本論文(DOI: 10.5281/zenodo.24098552)は、無断学習された特定の知的財産データを、AI事業者の全額コスト負担によって学習済みモデルのパラメータ空間から完全に消去させる「機械的忘却(マシン・アンラーニング)強制権」の法理・数理・暗号学的実装体系を確立するものである。推論時の表層的な出力フィルタリング(RLHF等による隠蔽)はモデル内部の重み空間における知財搾取状態を解消せず、情報理論的な忘却要件を満たさない。本研究では、64ビットGoldilocks素体(F_p, p = 2^64 - 2^32 + 1)上のQFS-zkVM、RTO P2.0、Q-SAFA、およびHuman-in-Command(HIC)統制下のPHOENIX_BRIDGE / PHOENIX_ROUTERハードウェア・インターロックを統合し、対象知財データの影響がモデル重みから完全に除去されたこと(またはクリーンルーム環境での再学習が完了したこと)をゼロ知識証明によって決定論的に検証するプロトコルを定式化する。さらに、消去・再学習・暗号学的監査にかかる計算および財務コストの一切を無断学習を行った企業側に強制負担させる責任原則と、SEC Rule 15c3-5およびデラウェア州会社法Caremark基準に基づく取締役会の監視義務違反責任を明確化する。