Five Filters in AI Dialogue and Multi-Turn Misreading Sliding

Multi-turn interactions with generative artificial intelligence can remain locally coherent while the relative importance of information shifts across the dialogue. This short note proposes an operational method for separating five turn-level filters: task formulation, wording, observable interpretation of user intent, adequacy of the response, and user comprehension. It distinguishes factual-error trajectories from trajectories in which otherwise relevant material gradually changes priority. A deterministic, assumed-parameter state-transition model illustrates how verification frequency and repair rates can alter the share of trajectories aligned with an initial task after twenty turns. With illustrative values, that share is 8.85 percent without audits, 89.49 percent with partial repairs after every turn, and 96.29 percent when both error classes receive an assumed eighty percent repair rate. These figures are mathematical outputs of stated assumptions. A one-factor sensitivity analysis is paired with an empirical protocol for coding dialogue logs, recording disagreement between raters, identifying shared verification sources, and assessing downstream repair. The protocol specifies observations that can test filter independence, the co-occurrence of error classes, repair effectiveness, and dependence on dialogue history. The contribution is a reproducible measurement proposal for subsequent comparative research.