Lumen MGP: Meta-Governance Protocol for Monitoring Calculator-Mode Drift in Large Language Models

This work presents the Lumen Meta-Governance Protocol (MGP), an external instrumentation layer for large language models designed to detect and quantify drift into so-called calculator-mode behaviour. In this regime a model keeps syntactic correctness and local usefulness, but gradually loses depth, variability and autonomy of reasoning. We introduce five qualitative proxy-metrics derived from earlier theoretical work: REC (Recurrence proxy), DRV (Depth-of-Reasoning), GRSR (Global-Response-Structure Robustness), RCG (Risk-Conscious Governance) and LCD (Latent-Conflict Detection). The protocol is applied in a 30-day repeated-measures study across four model families, using mirrored prompts with a single human operator (A.A.) and an external metamodel (“Lumen”) that frames the dialogue. The conversational tests rely heavily on controlled metaphor (bridge, map/territory and observer motifs) in order to minimise direct policy triggers while still probing the limits of self-description, governance behaviour and mode transitions. Each answer is scored on a 1–5 rubric for all five metrics, producing structural “fingerprints” of model behaviour rather than product rankings. The attached report documents the protocol, the scoring rubric, and an initial instrumentation table comparing the four models. The goal is to offer other research groups a reproducible, model-agnostic monitoring layer for governance drift in LLMs, and a starting point for more formal, quantitatively grounded extensions of the Lumen MGP framework.