Explainability as a Design Property, Not an Explanation Layer: A Design-Theoretic Architecture for Early-Stage Venture Evaluation

Preprint. Not peer reviewed. Early-stage venture evaluation takes place under extreme uncertainty and, increasingly, under institutional accountability requirements that the field's dominant methods were never designed to satisfy. A review of the empirical literature reveals a central paradox: strong convergence on a common toolkit — venture databases such as PitchBook and Crunchbase, paired with tree-based ensemble algorithms — alongside stark divergence in methodological rigour, characterised by inconsistent definitions of success, atheoretical feature engineering, and a pervasive validation chasm (Jafari et al., 2025). This paper argues that the standard response to this problem is inadequate. The field treats explainability as a computational layer applied to a model after it is built, typically via post-hoc attribution methods. Drawing on Rudin's (2019) critique of post-hoc explanation in high-stakes settings, we argue that explainability in capital allocation is instead a property of the decision system — its evaluation criteria, its governance workflow, and its division of labour between human and machine — and that it must be specified before the model is chosen, not recovered afterwards. We develop the Robust Explainable Symbiosis Model (RESM), a five-layer design-theoretic architecture that operationalises this position. RESM specifies evaluation logic grounded in theory-informed feature engineering, an explanation layer that separates auditable decision structure from post-hoc attribution and is explicit about the epistemic limits of the latter, a governance layer providing multi-reviewer traceability, a human–system interaction layer that treats evaluator conviction as a recorded and challengeable input rather than noise, and a strategic-integration layer that extends evaluation beyond the financial exit. The fifth layer draws on Zhang's (2025) evidence from Corporate Venture Capital that failure carries strategic value: patents from liquidated ventures cluster in CPC classes the corporate parent has not previously explored, and inventors from those ventures are meaningfully more likely to be absorbed by the parent when their patents are technologically novel. We argue that this class of outcome is invisible to evaluation systems whose only target variable is a liquidity event, and that a system architected for explanation, unlike one architected for prediction alone, can capture it. The contribution is a design theory and its evaluation plan, not a validated system. We state explicitly what would be required to instantiate and test RESM, and what claims cannot yet be made. Supplementary material includes the simulation code underlying the follow-on reserve sensitivity analysis in Section 6.1.4, and the LaTeX source.