What Shall We Call It: Naming, Epistemic Provenance, and the Structural Conditions of the Mythos Disclosure

The 27 March 2026 Anthropic data leak exposed two draft blog posts describing the same unreleased model under two different names: Claude Mythos (v1) and Claude Capybara (v2). The subtitle of the Capybara draft still read: “We have finished training a new AI model: Claude Mythos.” The organization had not resolved the model’s public identity at the time of accidental disclosure. This paper, the second in a three-paper series responding to the leak, argues that the unresolved naming is not a communications detail — it is an epistemic provenance failure.Applying the Emotional-Epistemic Provenance Layer (E-EPL) framework developed in Maijala (2026a), this paper identifies the Mythos/Capybara naming gap as a Layer 2 epistemic provenance failure at the institutional level: a condition in which an organization has built a capability of unprecedented self-described risk without completing the public reasoning chain about what that capability is, what it will be called, or how it will be positioned. The paper situates this failure within the structural conditions that produced it — competitive pressure from concurrent frontier model development, IPO-driven development velocity, and a governance architecture that did not require epistemic identity stabilization before external disclosure risk existed. The paper further argues that the Capybara tier decision — the structural determination that the existing Opus tier is insufficient to contain the new model’s capability level — is the most analytically significant disclosure in the leak, and that its near-total absence from public coverage reflects the same epistemic incompleteness it documents.Keywords: epistemic provenance, E-EPL, Claude Mythos, Capybara, naming, frontier model governance, AI safety, CBIL, development velocity, dual-use