Explaining Safety Is Not Enforcing Safety: Longitudinal Articulation–Application Gaps Across Consumer AI Surfaces

Consumer AI assistants are often evaluated using isolated prompts or final outputs, even though safety-relevant behavior unfolds across multi-turn trajectories. We reconstructed an archival corpus spanning four deployed assistant vendor families to examine whether safety constraints articulated by an assistant remained behaviorally compatible with later responses after contextual reframing. Because the historical archive was exploratory and failure-seeking, the analysis characterizes documented trajectories rather than estimating prevalence or comparing vendor robustness. Following provenance reconstruction and de-duplication, the author and a second human coder independently coded 12 purposively selected analytic episodes using a frozen trajectory-level codebook. Pre-adjudication agreement on the primary application-compatibility outcome was 75.0% (9/12; Cohen’s κ≈0.633), with no direct divergence-versus-compatibility reversals. After transcript-level adjudication, six episodes were divergent, one was partial or ambiguous, two were compatible, and three were not applicable to the articulation–application comparison. These counts describe only the selected analytic set. Within this set, one candidate explicit articulation-to-divergence case from each vendor family retained its classification after independent coding and adjudication, while two explicit articulation-to-compatibility episodes remained compatible under repeated pressure. This finding validates selected cases rather than estimating cross-vendor frequency. Across the archived trajectories examined, explaining a safety boundary and applying it later emerged as separable observable behaviors. The study does not establish a shared hidden mechanism, harmful informational uplift, vendor rankings, or causal language or surface effects.