<p>This article proposes a pragmatic-cognitive device for argumentation in mediation as a way of operationalising standards of reasonableness in real-time, non-ideal interaction. The problem addressed is not only how to evaluate argumentation after the event, but how to support argumentative quality while interaction unfolds under conditions of emotional intensity, asymmetry, and limited time for reflection. Mediation is treated as a revealing case because it requires managing disagreement through live argumentative activity without interrupting dialogue or replacing the parties’ own reasoning. The device is defined as a process-oriented and meta-argumentative structure that translates pragma-dialectical norms into situated communicative procedures. In its minimum form, it combines an aim, a set of normative constraints, a repertoire of communicative operations, and a processual result. Within this framework, selected NLP-derived tools are treated in a limited sense, not as a self-sufficient theory of cognition or communication, but as operational resources whose admissibility depends on their compatibility with pragma-dialectical norms and mediation principles. The article organises the device around four interdependent dimensions—normative, adaptive, structural, and relational—and illustrates their operation through a reconstructed family mediation case. Its main theoretical contribution lies in identifying an operational gap within existing argumentation theory and in proposing a meso-level, process-warranting mechanism for supporting discussability, examinability, and party agency in real-time interaction. A related contribution lies in clarifying responsiveness as the pragmatic counterpart of reasonableness under non-ideal conditions. The article closes by outlining implications for training, evaluation, and protocol design, while stressing the need for empirical validation.</p>