Document-to-Simulator Compilation: Toward an Executable Layer of Scientific Communication

Scientific communication rests on two layers: publication and data. This paper argues for a third layer: an executable layer that derives from a document itself a synthetic environment for structured exploration of its internal logic. The central thesis is that many structured knowledge artifacts can be treated as latent simulator specifications: they encode entities, variables, constraints, temporal structure, and outcome definitions that can support parameterized, constraint-enforcing, Monte Carlo-capable simulation environments without access to original data. The current implementation is semi-automated and domain-configuration-aware: automated extraction identifies candidate structure, while domain expertise is used to review constraints, plausibility criteria, and validation checks. A six-stage pipeline - artifact ingestion, structural extraction, schema construction, constraint enforcement, generative simulation, and validation - operationalizes this transformation. The framework is demonstrated through a concrete case study in which Vinton G. Cerf's discussion of NTP and UNIX timestamp overflow hazards is compiled into a running simulator with remediation scenarios, datasets, and figures. Additional cross-domain demonstrations include clinical cohort modeling, aerospace human reliability, robotics reinforcement learning, biosignal simulation, and narrative-to-data conversion. The practical implication is infrastructural: executable companions could enable methodological verification without data sharing, support open-science mandates in privacy-sensitive fields, and allow direct experimentation with the logic of published knowledge. The paper distinguishes executable reproducibility from scientific reproducibility and articulates limits, including synthetic-data fidelity, domain configuration, and the need for real-world validation in safety-critical applications.