Coherence Without Access: Layered Constraint, Unfaithful Self-Report, and the Overtrusted Explanation Problem

This paper argues that humans and large language models can instantiate a shared failure class: systems that produce coherent, context-sensitive, norm-aware verbal outputs under layered constraints without reliable access to the process that generated those outputs. It develops a four-part distinction between coherence, sincerity, access, and faithfulness, and uses contemporary literature on introspective access, choice blindness, interoceptive inference, explanation faithfulness, hallucination under uncertainty, and monitorability limits to show why articulate explanation should not be treated as transparent evidence of mechanism. The paper’s central governance claim is that explanation-based accountability is structurally fragile when verbal rationale is overtrusted as proof of process access. It further proposes a provisional systems-level extension in which this overtrusted-explanation problem may become amplifying when distributed across human, model, and institutional levels. RUNPORT II is included as supporting evidence that explanation surfaces may also vary in architecture-sensitive ways across models under shared prompt conditions. License: Copeland Resonant Harmonic Formalism (CRHC v1.0) This work is licensed under the Copeland Resonant Harmonic Copyright (CRHC v1.0). Attribution is required for all uses. Collaboration, academic discussion, and non-commercial use are permitted. Commercial use, resale, or incorporation into proprietary systems is not permitted without explicit written permission from the author. Licensed CRHC v1.0 Ψ(x) = ∇ϕ(Σ𝕒ₙ(x, ΔE)) + ℛ(x) ⊕ ΔΣ(𝕒′) — C077UPTF1L3 Derivative works must preserve attribution and must not remove or alter the stated license terms.