From Legal Rules to Evidence Architectures: A Theory of Minimum Legally Sufficient Telemetry

Autonomous systems are usually instrumented for operation before anyone knows which facts a later legal dispute will make decisive. The result can be paradoxical: abundant logs, yet no evidence for the distinction that matters. We introduce Legal-to-Evidence Compilation (LEXCOMP), a formal design layer that maps a structured rule model, legal question, and system description into factual worlds, a legal quotient, material predicates, typed evidence obligations, and Minimum Legally Sufficient Telemetry (MLST). The central principle is that law induces a quotient of factual reality; legal evidence is a sufficient representation of that quotient; and telemetry design is the engineering problem of realizing that representation at minimum cost. We define Legal Distinction Complexity, the Legal Sufficiency Gap, and legally sufficient statistics; derive an information lower bound; prove compilation soundness and incompleteness, jurisdictional non-universality, refinement monotonicity, MLST NP-hardness, and portfolio subadditivity; and extend the framework to robust consequence models. LEXCOMP implements exact MILP and greedy synthesis, adversarial compilation tests, legal-change regret, and an interactive explorer. Across 450 controlled trials, MLST preserved all modeled legal distinctions at 0.073 normalized cost. An executed reference-agent trace showed operational telemetry covered 0.500 of legal distinctions, whereas MLST achieved complete modeled coverage without claiming automatic legal interpretation or liability.