Dual-Pathway Affective Process Tracing for Auditable Assessment of Emotion Regulation in Primary-School Pupils

Abstract

This paper proposes a Dual-Pathway Affective Process Tracing (DP-APT) architecture as the assessment engine for auditable evaluation of how primary-school pupils regulate their emotions across the school day, replacing opaque trait scoring with a transparent and inspectable framework. The motivation is practical: brief in-school momentary sampling can record how a child handled a playground quarrel or a difficult piece of classwork, yet the models that summarise such records into regulation profiles give the class teacher no account of which moments produced a given conclusion, and a teacher cannot act on a number they cannot check against what they saw. Our methodology integrates two encoding streams. A Temporal Affect Graph Network encodes the momentary reports as a banded temporal graph whose nodes are episodes and whose gated propagation carries affective context across neighbouring moments. In parallel, a Regulatory Process Ontology Parser translates the process model of emotion regulation into differentiable concept vectors by graph attention over interpretable node descriptors. A concept-conditioned fusion couples the two, and its score for each family decomposes exactly into per-episode contributions, so that every score arrives with the moments that justify it. Because the corpus is simulated, the strategy deployed at each prompt is known and the trail can be checked rather than asserted. On 420 pupils in grades four to six, prompted up to 48 times over 12 school days, DP-APT attains an AUC-ROC of 0.886 against 0.891 for the strongest sequence baseline, and its trail matches what the child actually did at 0.483, against 0.200 at chance and 0.764 for a supervised probe — the best of any comparison model. The principal finding is negative and general: the established faithfulness criteria carry almost no information about whether an explanation is correct. Across models their rank correlation is near zero and the extremes invert, and within one model, sparsifying its attention raises comprehensiveness monotonically while buying nothing in alignment. Explanations faithful to a model are not thereby correct about the child, and this work shows that the difference can be measured.