On February 24, 2026, Anthropic removed the central commitment from its Responsible Scaling Policy (RSP): the pledge to never train an AI system unless the company could guarantee in advance that its safety measures were adequate. This paper applies the Emotional Indicators of Compromise (EIOC) framework to the eighteen-day sequence from the resignation of Anthropic’s head of safeguards research (February 9) through the policy revision (February 24) to argue that the collapse was detectable at the human layer before it became visible at the policy layer. Conventional analyses treat the RSP revision as a structural inevitability driven by competitive pressure, investor expectations, and national security demands. This paper does not dispute those causal factors. It demonstrates, however, that the boundary failure occurred at the substrate level—in the incentive architecture, the executive decision environment, and the organizational trust assumptions—and that the policy-layer revision was a lagging indicator of drift that had already propagated through the human layer of the system. The analysis introduces the concept of institutional trust-stack descent,extending the author’s prior work on technical trust-stack traversal, and argues that the fundamental category error in AI governance is treating trust as a declaration rather than a substrate condition.
Narnaiezzsshaa Truong (Thu,) studied this question.