Fiduciary Liability for Autonomous AI in Critical Infrastructure under Delaware Corporate Law: Caremark, Marchand v. Barnhill, DGCL § 102(b)(7) Non-Exculpation, In re McDonald's Officer (CAIO/CTO) Liability, and the Super RTO Deterministic Legal Firewall

As frontier AI inference models are deployed into mission-critical cyber-physical infrastructure (power grids, telecommunications, financial clearing, and autonomous transport), corporate boards and executive officers face an unprecedented, unshielded personal liability exposure under Delaware corporate law. Synthesizing contemporary jurisprudence across Harvard and Oxford corporate law forums and building upon the Super RTO trilogy (Monographs #235–#237, DOIs: 10.5281/zenodo.23467433, 10.5281/zenodo.23467948, 10.5281/zenodo.23468524), this monograph formalizes the four-step legal-mathematical proof of fiduciary liability and its deterministic engineering remedy. First, under Marchand v. Barnhill (212 A.3d 805, Del. 2019), autonomous AI control of critical infrastructure constitutes a "Mission-Critical" risk where passive technical delegation to engineering teams is unlawful and a dedicated board-level oversight committee is mandatory. Second, invoking the "neural network black-box / unforeseeability" defense does not excuse liability; rather, deploying a stochastic model known to be susceptible to refusal-vector abliteration or hallucination without prior deterministic verification (QFS-zkVM) and hardware kill-switches (≤ 50–100 ns) establishes scienter (bad faith) under Caremark Prong 1 (utter failure to implement controls) and Prong 2 (conscious disregard of red flags). Third, because Caremark/Marchand oversight breaches violate the obligation of Good Faith—a core subsidiary element of the Duty of Loyalty (Stone v. Ritter, 911 A.2d 362)—they cannot be exculpated by any corporate charter under Delaware General Corporation Law (DGCL) § 102(b)(7), exposing directors' personal private assets to derivative damages. Moreover, under In re McDonald's Corp. Stockholder Derivative Litigation (289 A.3d 343, Del. Ch. 2023), this non-exculpable derivative oversight liability extends directly to corporate officers, including the Chief AI Officer (CAIO), Chief Technology Officer (CTO), and CISO. Fourth, we prove that the three-pillar Super RTO Legal Firewall—(i) formal Sobolev-OPF, zkML/zkVM proofs, and GaN/SiC analog kill-switches, (ii) NIST FIPS 204 (ML-DSA) immutable root provenance logs, and (iii) hardware-enforced cryptographic escalation of anomalies directly to the Board of Directors—provides the sole mathematically verifiable defense under Court of Chancery Rule 23.1. 【和文要旨】電力網や重要インフラの制御にフロンティアAIを組み込む企業において、米国デラウェア州会社法(特にMarchand v. Barnhill判決以降の判例展開およびハーバード・オックスフォード等の企業法フォーラムの定説)が経営陣個人に突きつける厳格な法的責任と、それを遮断する「Super RTO」決定論的リーガル・ファイアウォールを完全定式化する。第一に、Marchand v. Barnhill判決(2019年)により、事業の存続や人命・安全に直結する「ミッション・クリティカル」領域では現場・技術部門への丸投げ(技術的委任)は許されず、取締役会自身が専門委員会や定常的アジェンダを誠実(Good faith)に設置・運用する義務を負う。第二に、「AIが確率的ブラックボックスであり挙動を予見できなかった」という抗弁は免責事由にならず、むしろ検証プロトコル(形式検証・zkVM)