Role Boundaries in the Type System: Accountable AI Assistance under India's POSH Act

A statute can say which decisions belong to which people. Software placed inside that procedure has to respect the assignment, and doing the job well is no defence: an assistant that writes an accurate, well-argued recommendation has still overstepped if the law gave that call to someone else. We argue that a boundary of this kind belongs in the type system, where a schema turns away an out-of-role request before the model runs, and that differential privacy can be turned around in the same setting, from a way of withholding information into the thing that makes a release lawful. India's Prevention of Sexual Harassment (POSH) Act, 2013 is where we test both claims. It fixes deadlines, names who decides what, and forbids disclosure. A decade after it passed, 219 of the 300 largest companies on India's National Stock Exchange report zero cases a year. We trace the gap to three things the Act assumes and nobody builds: deadline tracking, a record that cannot be quietly edited, and any visible sign to employees that the system runs. ReportItNow supplies all three without touching the Internal Complaints Committee's power to decide under \S\,11. The statutory procedure becomes a state machine. Six role-scoped assistants sit on top of it behind typed schemas. A two-tier dashboard publishes annual aggregates under the confidentiality proviso in \S\,16. We report a threat model with an added legal category, an adversarial test layer aimed at the boundaries we defend least well, and a Monte Carlo simulation showing that at $\varepsilon = 1.5$ per fiscal year the published figures carry real information only from about 60 cases upward. Below that a release is honest and uninformative.