Large language models (LLMs) are replaced by new versions at short intervals, often without users being informed of how these updates affect model behavior. Established evaluations largely measure what models can do; far less attention has been paid to how they frame contested or normatively charged questions, and to whether this framing behavior changes across versions. This study introduces an empirical case study and methodological demonstration of Prompt Echo Tracking (PET), a behavioral auditing protocol for detecting framing-related behavioral drift across LLM versions. PET uses a four-phase design with two baseline measurements (A1, A2), a contrast condition (B), and a post-measurement under restored neutral conditions (C). This design separates conditional reactivity to an active framing intervention from persistent effects that remain after the intervention is removed. We apply PET to a real-world version transition between Claude Opus 4.7 and 4.8, using five German-language climate prompts and 640 observations per model, with responses independently recoded by three coder families (GPT-4o, Gemini, Mistral). The headline finding is a reproducible, coder-robust shift in framing-related behavior between the two versions under otherwise identical conditions: under the same evaluative framing, Opus 4.7 moved toward a directive, normatively committed stance on climate responsibility, whereas Opus 4.8 remained comparatively multi-perspectival. The drift is a redistribution of framing responsiveness across prompt types rather than a global increase or decrease. No persistent effects were detected once the framing intervention was removed, which locates the shift between model versions rather than within conversations. The findings suggest that framing-related behavior can shift across model versions in ways that capability-oriented benchmarks would not necessarily detect; PET offers a reproducible protocol for making such shifts observable. The evidence rests primarily on a single model pair from one provider and is to be understood as a case study; a supplementary check on a second, cross-tier model pair from the same provider is reported in Section 3.4. This is an author-formatted preprint. A German-language preprint of the underlying concept paper is available on SSOAR (see related identifiers below). Peer-reviewed submission is planned separately.
Denis Pijetlovic (Sat,) studied this question.