From Talking to Blackbox to Governance: Constitution Mining for LLM Agents

LLMs now operate as decision mediators inside tool-using agents, where outputs can become external side effects. In this regime, the core problem is no longer only correctness, but governability: the system must justify when it is permitted to execute, when it must legitimately hesitate, and when it must refuse. This paper frames the work in the “talking to a black box” stance: rather than claiming internal access or total transparency, we treat black-box interaction as a disciplined dialogue that converts opacity into stable, testable artifacts. Building on a conservation view of interpretability — where a decision carries an interpretable share and a residual opaque share, and explanation is the iterative transfer from opacity to clarity — we introduce Constitution Mining, a black-box methodology that extracts permission logic as an executable constitution. A constitution is represented as K = (R, Π), where R is a finite set of typed if/then rules and Π resolves conflicts either via precedence or via a true partition over a declared input domain. The method enforces semantic binding, schema discipline, firing separation (Satisfied vs Considered), and partition validation (coverage and exclusivity), producing replayable artifacts (rules, receipts, drift boundaries) suitable for audit and regression under model updates — without asserting mechanistic interpretability.