governing access is precise work — who has access to what, who approved it, what changed. that precision usually lives behind a dense interface. this lab asks whether we can compress it into natural language, without losing the safety.
the problem
“who has access to payroll?” is an easy sentence and a hard question. the gap between the way people ask and the way systems answer is where mistakes and over-privilege hide. we want the speed of speech without the sloppiness of speech.
deconstruction
understand before you change
the first cut is read-only: a question returns an answer, scoped to what you can already see. nothing is granted, nothing is changed. comprehension before mutation.
propose, never act
the second cut separates suggestion from execution: the system can propose a change and a human confirms it. it never acts on its own. the decision stays with the person.
structure over prose
an answer is never a free-form paragraph — it is a bounded result: a conclusion, the things that matter, and at most one next step. bounded output is an audit trail, not a conversation.
fail closed
when unsure, the system asks rather than guesses. the cost of a wrong answer in access governance is too high to leave to confidence.
tensions
- expressiveness versus safety — language is fluid, governance is exact.
- “speech to code” versus “speech to command” — generating code is open-ended; emitting a bounded command is safe. we want the latter.
where this lands
an assistant that understands a question, drafts the answer, and waits for a human.