Graph RAG and the unofficial rules

There is a version of every institution that exists in the policy manual, and there is the version that actually closes the quarter. The gap between them is the unofficial rules: the threshold under which a regional lead does not escalate, the counterparty that always gets an extra day, the account that is reconciled by the same analyst because only she knows why it never balances. None of this is in the data lake. Some of it is in the dark-matter data lake. Most of it is in the heads of a small number of people who are, statistically, about to leave.

Retrieval-augmented generation over documents retrieves the official rules and is, for that reason, a very polite way to be wrong. Graph RAG retrieves structure. It models the institution as entities, relationships, exceptions, and the provenance of each exception, then lets the agentic runtime traverse that graph at inference time. When a coverage officer asks why a limit was set, the answer is not a paragraph from a policy. It is the chain of approvals, the person who made the call, the precedent they cited, and the three times the rule was quietly bent since. Institutional knowledge becomes a path rather than a paragraph.

The construction problem is that unofficial rules are never declared. They are observed. The fabric learns them from the residual: the approval that happened faster than policy allows, the escalation that did not occur, the reconciliation that lands on the same immaterial variance every month. Causal inference separates habit from noise. Constitutional AI decides which habits are load-bearing and which are liabilities that should not be preserved. The result is a graph annotated with confidence, provenance, and a flag that distinguishes rules the institution would defend from rules it would prefer the regulator not see.

This is where the last managing director matters. When a senior person departs, the loss is rarely the decisions they made. It is the decisions they prevented, and the reasons that were never articulated because nobody asked. Graph RAG, populated continuously by the multi-agent runtime, has already asked. The successor inherits a traversable record of judgment, not a handover deck. Federated learning ensures the graph never leaves the sovereignty boundary, and zero-knowledge proofs allow the successor to verify the record's integrity without revealing the individual conversations it was inferred from.

There is an obvious governance question, and we treat it as the primary design constraint rather than a disclaimer. An institution does not want every unofficial rule preserved. Some are exposure. The graph therefore carries an explicit disposition for each inferred rule: codify, tolerate, retire, or refer to compliance. Explainable AI is the mechanism by which the change-control board reviews those dispositions, and PPoE attestation records the decision so that a later auditor can confirm that a rule was retired, not merely relabeled. The fabric remembers, and it remembers on the record.

Institutions that adopt Graph RAG for the unofficial rules tend to describe the same first surprise. The graph is larger than they expected, and the density is highest in the parts of the business they believed were most process-driven. This is not a failure of the institution. It is what a functioning institution looks like once its actual operating thesis is made legible. The last managing director is welcome to retire. The rules stay, with their provenance, their confidence, and their disposition, inside the building where they were made.