Knowledge graph: how the company brain holds people, structure and history as evidence

Most questions people ask about an organisation are questions about relationships. Who approves this? Who does she report to? Which group gives him access to that system? A document store can find text that mentions a name, but it cannot follow a chain from a person to a group to a permission, and it cannot tell you that the chain changed on Tuesday.

A graph makes those relationships first-class. Each one is stored once, with its source, its status and its history, and can be followed in either direction. Documents still matter, and the brain is designed to hold them too, with their access lists. But the structure that decides who may see and do what lives in the graph.

The graph is insert-only. When a fact changes, the old version is not overwritten: a new version is recorded with the time it became true, and the old one keeps the time it stopped being true. Nothing is quietly edited in place, so the record of how the organisation looked last year is still there next year.

That makes as-of reads possible. Ask who was in the finance approvers group last March and the graph answers from what was true then, not from today’s membership. Reviews, audits and investigations need exactly this: a decision judged against the organisation as it stood when the decision was made, after people have moved teams and groups have been renamed.

Coverage is reported as a measurement: which sources are connected, how fresh each one is, and how complete the picture is relative to what the brain can see. It is a measure, never a claim to understand everything. The local API serves coverage today; the views that draw it as charts and maps are on the roadmap.

Provenance answers a different question: where did this fact come from, and why did it change? Every fact already carries its evidence and its history. The endpoints that walk that trail for a single fact, the provenance and timeline reads, are in progress and not finished, so explaining a change today means reading the history directly rather than asking for an explanation.

The appliance records its own actions in an audit log in which every entry is chained to the one before it by a hash. Checkpoints over the chain are signed. If an entry is altered or removed after the fact, the chain no longer matches its checkpoint, and the tampering is detectable by anyone who checks.

The audit log is readable through the local API, scoped by token like everything else. It records what the appliance did, so an organisation can show what happened and when. It is one part of the evidence an organisation keeps; it does not by itself decide whether any particular use of the brain was appropriate. That judgement stays with people.

An internal auditor is reviewing why a contractor could approve supplier invoices in the spring. Today the contractor has left and the approvers group has a new name. The auditor asks the graph, as of the date of the first approval, who was in the group and how the contractor came to be there.

The graph answers from history. The contractor’s account was a member of a nested group that sat inside the approvers group, so transitive membership made the contractor an approver. The membership edge is observed, from the directory, with the times it was first and last seen. The contractor’s manager on that date is there too. What the graph cannot say is why the nesting was set up: that decision was never in a connected source, so the answer is unknown, and the auditor asks a person.

A vector store finds passages that are similar to a question. A knowledge base collects documents for people or a model to read. Both are useful, and the brain is designed to include permission-aware search over content as well. Neither keeps relationships with history and evidence. The graph is the part that does, which is why it sits at the centre.

The limits are real. Today the graph holds people, accounts, groups, org units, service identities and devices from four directory sources. It does not yet hold systems, services, ownership, documents or decisions. It knows nothing a connected source has not told it, and it will say unknown far more often than a product that guesses. That is intentional.