PAD-S/CSA as a candidate shared representation layer for computational psychotherapy: minimal architecture and a...

Psychotherapy schools often describe overlapping process phenomena in non-interoperable vocabularies. This pluralism is clinically valuable but computationally costly: datasets become difficult to compare, clinically load-bearing distinctions are collapsed into convenience labels, and artificial intelligence (AI) systems inherit annotation schemes rather than a clinically interpretable intermediate representation. Building on the Perceive-Assess-Dose-Safeguard (PAD-S) framework and the Conflict-Square Algorithm (CSA), this theory article asks a narrower question than the prior PAD-S and CSA papers: can the same variables be formulated as a candidate shared representation layer between heterogeneous observation models and school-specific intervention policies? The proposed layer projects a high-dimensional biopsychosocial state into four clinically observable process coordinates-defensive/avoidant organization (DEF), anxiety/arousal and tolerance (ANX), progression toward direct experience and action (PRO), and self-attack/shame processes (SUP)-plus a safety threshold that constrains admissible intervention intensity. The contribution is architectural rather than empirical: it isolates the representational role from earlier decision-grammar and transcript-coding roles; clarifies the distinction between observations, representation, and policy; specifies a minimal falsifiable family of state-transition models; illustrates translation across four pragmatic therapy families; and defines a staged validation order from reliability and function linkage to transcript-level predictive operationalization and only then sparse equation discovery. The framework should therefore be read as a candidate shared representation layer for computational psychotherapy and computational psychiatry rather than as a therapy protocol, a fitted predictive model, a complete generative theory, or an autonomous decision system. No new dataset, fitted classifier, transcript-level predictive result, or discovered equation is reported here. The article aims instead to state what would count for or against PAD-S/CSA as a clinically interpretable interface for later empirical modeling.