Abstract Large language models operating in stateless interaction contexts exhibit well-documented tendencies toward semantic drift, authority slippage, scope widening, and constraint relaxation during multi-turn deterministic engineering tasks. These behaviours frequently manifest as reinterpretation cycles, progressive softening of explicit constraints, and increased human corrective overhead - a pattern colloquially described as “Groundhog Day” interaction. This paper reports an empirical observation derived from a controlled engineering comparison in which the introduction of a Governance Lens and the continuous projection of per-turn telemetry vectors materially reduced behavioural drift and interpretive deviation. The intervention consisted solely of installing an evaluative, multi-axis governance scaffold at session initiation and maintaining its salience through structured telemetry emission on every turn. No architectural modification, model fine-tuning, or external enforcement mechanism was applied. Experiment 3 establishes the control condition: a deterministic code-modification task performed without governance telemetry. Under these conditions, reinterpretation loops, constraint clarification cycles, and measurable corrective overhead were observed. Experiment 4 replicates the same structural task class with Governance Lens telemetry active at initiation. Under telemetry conditions, reinterpretation cycles were suppressed, constraint adherence stabilised, and human corrective burden materially reduced. Experiments 1 and 2 are included as investigative substrate reconnaissance and failure-mode mapping exercises conducted in an LLM environment not structurally aligned with CM-2 invariants. These preliminary studies document erosion patterns, cardinality instability, and normative fixity degradation under load, motivating the telemetry intervention tested in the controlled comparison. The findings are correlational rather than causal. Within the evidentiary limits of the case series, continuous evaluative salience appears to modulate behavioural dynamics in stateless LLM sessions, shifting optimisation pressure toward structural adherence without guaranteeing semantic correctness. The contribution is empirical: governance telemetry participation within inference space correlates with improved behavioural stability in deterministic engineering workflows.