Post-publication note (2026-05-23T06:36Z): Subsequent to the completion of this study, later interaction sessions with ChatGPT indicate that interactive UTC wall-clock time may once again be accessible under at least some deployment conditions. The observations documented in this paper nevertheless remain historically accurate for the observation interval analysed herein. The author has not re-evaluated whether the apparent restoration of interactive time access is sufficient to re-establish full CM-2 temporal compliance, including Self-Host Bootstrap serialisation workflows, deterministic Epistemic Object envelope generation, immutable `created_at` assignment, or Reference Object Collection (ROC) replay semantics. Accordingly, the architectural conclusions of this paper remain unchanged: CM-2 protocol operability depends on stable exposure of platform-level temporal affordances, and changes in platform interface behaviour can directly affect the implementability of higher-level distributed cognition protocols. Abstract This paper documents an observed behavioural regression in ChatGPT affecting the generation of artefacts required for the CM-2 Cognitive Memoisation protocol. CM-2 defines normative invariants for Epistemic Objects (EO) and Epistemic Attributes (EA), including the requirement that each artefact contain a platform-generated wall-clock timestamp (created_at) expressed in ISO-8601 format. These timestamps serve as temporal anchors for epistemic identity, provenance tracking, and Reference Object Collection (ROC) cohort discrimination. Historical corpus evidence—including TMLDUMP and XDUMP artefacts produced during earlier ChatGPT sessions—demonstrates consistent emission of timestamps closely aligned with session chronology. In recent sessions, however, ChatGPT explicitly refuses or is unable to produce wall-clock timestamps, despite the hosting platform clearly maintaining message-level time metadata. This behavioural change prevents compliant EO serialisation and disrupts several CM-2 mechanisms, including ROC cohort formation, checkpoint recovery, and deterministic temporal provenance. Using previously published corpus artefacts together with newly recorded interaction transcripts, this study analyses the regression as a platform-interface change rather than a modification of CM-2 protocol semantics. The findings highlight a structural dependency between governance protocols and the capabilities exposed by LLM platforms. The case illustrates how platform behaviour can directly impact the implementability of higher-level knowledge-governance architectures and underscores the need for stable interface guarantees when protocols rely on platform-generated temporal anchors.