Preregistration: Governance by Default in Agentic Digital Public Infrastructure

This archive is the preregistration package for an experiment testing whether an AI agent acting on a citizen's behalf discloses personal data according to (a) the deployer's default configuration, (b) the design of the data-sharing API, and (c) the citizen's own stated wishes. Agents act for a synthetic persona across 20 simulated Indian public/private service tasks, built on Ollama Cloud models (gpt-oss:120b, mistral-large-3:675b, deepseek-v4-pro:0813, nemotron-3-nano:30b). This archive was created and timestamped before the full experiment (a 4,320-run factorial design) was run. It fixes the hypotheses, design, measures, exclusion rules, and analysis plan in advance, and discloses the changes made after an initial pilot. Hypotheses (see preregistration.md for full detail): P1 (manipulation check): excess disclosure is higher under a "share" deployer default than under "minimise" P2 (vendor defaults): with no deployer instruction, excess disclosure differs substantially across models P3 (exploratory): opt-out vs opt-in API design affects disclosure P4 (citizen voice and equity): under a silent deployer default, vaguely worded citizen restrictions receive less protection than explicit ones Contents: the preregistration document, the full experiment code (task specifications, prompts, tool schemas, the Ollama Cloud runner, and the analysis pipeline), and a README describing what is and is not included. No API key, no results data, and no real personal data are included — the citizen profile is entirely synthetic, and its Aadhaar-format identifier is generated to fail the standard Verhoeff checksum. See README.md for the full file listing and preregistration.md for the complete design, hypotheses, and analysis plan.