Disagreement Is the Signal

Every governance system in production today was built to declare truth. Agents have arrived to test that assumption at machine speed, and the verdict is unkind.

An agent executes a write against a Salesforce customer record. The column is tagged PII in the data catalog, untagged in the source schema, and the IdP shows the underlying principal acting on behalf of a human who is on vacation. The action proceeds. A regulator asks, six months later, who approved it. The honest answer is that nobody did. A system declared the policy. A system declared the tag. A system declared the identity. Each was technically correct in isolation, and the composition of their declarations produced a transaction nobody would have approved if asked.

This is the failure mode of declarative governance. It is not a corner case. It is the median outcome the moment execution leaves human hands.

The thirty-year frame

Enterprise governance has always been a declarative discipline. Master Data Management declared the canonical record. Role-Based Access Control declared the permitted action. Data catalogs declared the classification. Each system pronounced a fact and asked the rest of the stack to behave accordingly. For three decades this worked acceptably, not because the declarations were correct, but because the executors were human. People noticed when the catalog was wrong. People overrode the permission. People escalated the ambiguous case. The governance system was a starting point for human judgment, not a replacement for it.

The reason MDM never delivered on its promise — and the reason every analyst rated it a perennial disappointment — was not that the technology was bad. It was that the premise was wrong. No human committee can maintain a canonical record at the speed enterprise data actually changes. The committee is always behind. The record is always stale. The humans downstream knew this and worked around it, which is why the projects "failed" in spirit even when they "succeeded" on paper.

Agentic execution removes the human safety net. The catalog being out of date is no longer a paper cut. It is a regulatory event.

What changed

Three shifts have arrived together, and they are forcing the issue.

The first is speed. An agent does not wait to verify. It acts on what the governance system tells it, at the latency of an API call. There is no afternoon to think about it, no Slack to a colleague, no email to legal. The declaration is the decision.

The second is volume. A single principal can now initiate thousands of cross-system actions per hour. The exception-handling discipline that governance teams relied on — find the strange case, route it to a human — does not survive contact with that volume. Every action becomes the strange case.

The third is composition. A modern action touches five or seven systems before it completes. The catalog has one view of the entity. The source system has another. The IdP has a third. Lineage has a fourth. Token usage history has a fifth. Each of those sources is correct about its own slice of reality and uninformed about the others. The composition is where the governance question actually lives, and it is a question no declarative system was designed to answer.

A declaration is a verdict from one witness. The agentic era requires a deliberation across many.

The principle: Computed Confidence

The answer is not better declarations. It is the abandonment of declarative governance as the primary control. The new primitive is a confidence score, computed fresh at the moment of decision, by adjudicating across every available source and treating their agreement as the signal.

Call this Computed Confidence Governance. It has three properties that distinguish it from the architectures it replaces.

First, it is multi-source by construction. It does not pick a winner. The source system, the catalog, the IdP, the lineage graph, the token usage history, the policy engine — all of them are witnesses. None of them is canonical.

Second, it treats disagreement as information rather than as an exception. When the catalog says a column is PII and the source schema says it is not, that disagreement is not a failure to be reconciled offline. It is a real-time input. Confidence drops. Thresholds trigger. Routing changes.

Third, it produces an output that is itself probabilistic. Not "allow" or "deny," but "allow at confidence 0.87 under these conditions" or "route to human at confidence 0.62 because of disagreement between sources A and C." Downstream policy can be expressed against that probability, which means humans can finally specify the tolerance they actually have rather than the binary they were forced into.

Here is the shape of a single decision, in the form a governance plane would emit:

The catalog being wrong does not break the system. It lowers the confidence and routes accordingly. The governance plane has learned something — a source is unreliable, an owner needs to be notified, a remediation is queued — without halting the transaction or losing the audit trail. Disagreement is metabolized.

What this means for buyers and builders

For builders, the implication is that the catalog category and the MDM category are not the seat of governance anymore. They become substrate. A catalog is an excellent witness. MDM, where it exists, is another witness. Neither is the control plane. The control plane is the adjudicator above them, and it is a different product with a different buyer.

For buyers, the question shifts. The old question was "what is our master record?" The new question is "what is our confidence threshold, and who set it?" The CFO, the General Counsel, and the Chief Data Officer are about to discover that the threshold is the artifact of governance that actually has regulatory weight. Whoever owns the threshold owns the liability. Whoever owns the liability owns the budget.

For regulators, the implication is more interesting still. A declared policy is auditable but useless under speed. A computed confidence score, logged per transaction, is both auditable and operationally meaningful. The next generation of compliance evidence will not be policy documents. It will be confidence distributions, threshold settings, and disagreement logs.

The closing observation

The catalogs are not wrong. The IdPs are not wrong. The MDM systems are not wrong. They were all built for a world where a human executor stood between their declarations and the outcome. That world is ending. The control plane that replaces it must be built around the assumption that its witnesses will disagree, and that the disagreement is not a bug.

Disagreement is not the failure mode of federated governance. It is the feature.