The architecture of algorithmic fairness: bridging the gap between formal metrics and substantive justice

Abstract The dominant paradigm for judging algorithmic systems relies on statistical parity, a method that offers mathematical precision but proves fragile in the face of the lived complexity of injustice. While bias-mitigation techniques advance, they remain constrained by narrow formalisms that ignore the historical and structural conditions sustaining inequality. Consequently, algorithmic governance cannot be treated solely as technical optimisation. Against this unsettled background, this paper advocates viewing algorithms as socio-technical systems requiring substantive, not just formal, fairness. This study introduces the Substantive Fairness in Practice (SFIP) Framework, a normative-theoretical approach grounded in political philosophy and deliberative democracy. SFIP integrates structural analysis, power mapping, and stakeholder deliberation across five stages. By operationalising critical theory into a five-stage lifecycle, the framework offers scaffolding to translate substantive justice from an abstract ideal into a concrete engineering practice. Using a university admissions scenario as a heuristic, this study illustrates how SFIP operationalises accountability, making visible the power dynamics and historical structures that purely mathematical approaches do not address.