Interpretability Without Full Recoverability: Identifiability Limits and a Reportability Framework for Withdrawal-Side Parameters in the Continuous Moral–Social World Model

Mechanistically expressive computational models are often read as though plausible behavior were sufficient to justify parameter-level interpretation. In social computational psychiatry, that move is especially risky because several latent processes can generate similar interaction patterns while remaining weakly constrained by the observation model. We examine this problem in the Continuous Moral–Social World Model (CMS-WM), a three-layer simulator coupling a phenotype state, a world state, and a state-biased softmax policy over cooperation, defection, and withdrawal. The focus is the withdrawal-side policy decomposition, which contains ambiguity-linked, friction-linked, pressure-linked, and threshold-linked components. The study was designed as an identifiability and reportability analysis rather than a performance benchmark. We first established the structural ceiling for inference: the policy admits an exact scaling symmetry, so softmax temperature is structurally non-identifiable and only temperature-normalized effective parameters are legitimate inferential targets. We then conducted a staged audit of practical recoverability under a summary-based observation model, testing whether weak recovery could be attributed to limited optimization budget, compressed summaries, poor parameterization, inadequate excitation, low-dimensional tradeoff structure, reduced-model pruning, or targeted feature redesign. Across these analyses, withdrawal-relevant behavioral regimes remained comparatively stable, whereas withdrawal-side parameter recoverability remained sharply uneven. Only the ambiguity-linked withdrawal term survived a conservative quality gate across the tested configurations, and even then only as a cautiously interpretable quantity within the present simulator, scenario batteries, and summary-based inference pipeline. The friction-linked withdrawal term remained informative but configuration-dependent; the pressure-linked withdrawal term and withdrawal threshold did not support substantive interpretation in the current model-observation pairing. The contribution is therefore narrower, but more defensible, than a full parameter decomposition. CMS-WM supports robust regime-level interpretation of withdrawal behavior together with a limited parameter-level reportability framework. More broadly, the results show that in mechanistically expressive social decision models, regime interpretability may be more stable than parameter recoverability, and that this asymmetry should be treated as a primary scientific result rather than a technical afterthought.