The Administrative Ground Truth Deficit (AGTD)

Pakistan’s National Artificial Intelligence Policy 2025 commits the state to a deployment-first trajectory, embedding artificial intelligence across governance, education, health, agriculture, taxation, identity and public service delivery. The national conversation around this commitment measures readiness almost entirel as a function of inputs: skills trained, infrastructure built, funds allocated and regulations drafted. This paper argues that the dominant readiness vocabulary omits the single precondition on which every downstream deployment silently depends, namely whether the records and reality the system acts upon are reliable enough to make the system’s outputs valid in the Pakistani context at all. We introduce the Administrative Ground Truth Deficit (AGTD), a composite, deployment-level measurementframework that quantifies the divergence between an artificial intelligence system and the verifiable reality ofthe population it governs. AGTD organises this divergence across three substrates and seven measurablecomponents. The Record Substrate captures the integrity of the data the state holds, through RecordFragmentation, Institutional Data Drift and Verification Dependency. The Model Substrate captures theintegrity of what the model itself brings, through Training Representation Deficit and Evaluation LocalityDeficit. The Accountability Substrate captures the capacity of humans and citizens to detect and reverse error,through Contestability Weakness and Automated Error Amplification. The framework’s theoretical core is the manufactured confidence mechanism: in low-trust, paper-digitalenvironments a high deficit does not announce itself as visible failure but converts into confident, scaled,silent error, the most dangerous failure mode for a state. AGTD scores each component from 0 to 100,combines them through a weakest-link cascade rather than a simple mean, and returns a deployment risk band.We argue that AGTD should function as a mandatory pre-deployment audit within the regulatory sandbox theNational AI Policy already establishes. The contribution reframes the national question from “Can we adoptartificial intelligence?” to “Is our administrative truth strong enough to be automated?” Keywords: Administrative Ground Truth Deficit; AI Readiness Debt; Public-Sector AI; AI GovernancePakistan; Data Integrity; Citizen Contestability; Institutional Data Drift; Automated Administrative Harm;Manufactured Confidence; Low-Resource AI Governance

Перевод пока недоступен