Ask your company: questions in plain language, answers with their evidence

An answer from the brain is more than a sentence. It carries the sources it rests on, the evidence status of each fact, and how long ago that evidence was seen. Who leads the data platform team is a question with an answer; so is how sure the brain is about it, and since when.

That shape is deliberate. People act on answers. An answer that rests on a verified fact can be used directly. One that rests on an inferred or stale fact should prompt a check. One that comes back unknown tells you to ask a person, and the brain should say so plainly rather than produce something that only sounds right.

Who does the head of procurement report to, and who did she report to when the new supplier was approved? The brain can answer both today through the local API: the current reporting line, and an as-of read at the approval date. Each fact comes with its status, usually observed from the directory, and the time it was last seen.

Who can approve purchase orders above the threshold? The brain can tell you who is in the approvers group, including through nested groups. It cannot tell you that this group is the one that actually grants the approval right in the finance system, because no finance system is connected. The answer gives the membership and says the link to the permission is unknown.

What does our travel policy say about business-class flights? This needs search over documents, which is in progress. Today the brain cannot answer it. When document search is finished, the answer would cite the policy, and only for people who are able to read that policy in its source system.

Who owns the customer notifications service? Ownership needs code, CI/CD and ticket collectors, which are on the roadmap. Today the brain says unknown, and can list the people in any team the directory describes. It does not guess an owner from the name of a team or a group.

Cortex, the Swfte desktop app, is the designed front door for asking the brain. Today Cortex has its own knowledge bases with retrieval on the device, MCP tools, and local models through Ollama and LM Studio. It does not read the company brain yet. The wiring from Cortex, and from Nexus and Studio, into the brain is on the roadmap.

Until then, the local API is how the brain is asked: structured reads for people, groups, the graph, coverage and audit, scoped by token. That is useful today for internal tools and scripts, and it is the same foundation that the plain-language surfaces are designed to sit on, with the same access rules underneath.

When two sources disagree, for example one directory names a different manager from another, the fact is disputed and the answer says so, with both sides and their sources. It does not pick the more convenient one. When a fact has not been seen again within its freshness window, it is stale, and the answer shows its age.

This makes answers less smooth than a chatbot’s. That is the point. A confident answer built on a stale reporting line can send an approval to someone who left in the spring. Showing the status lets the reader decide whether to act or to check, and gives the people who own the source a reason to fix it.

The brain only knows what connected sources have told it, and today those are directory sources. Plain-language asking, document search and questions about systems and ownership are not available yet. Coverage is a measure, never a claim to understand everything, and an answer of unknown is a correct answer when the evidence is missing.

The brain does not decide whether an answer is appropriate to act on: the reader, or the policy of the workflow that asked, does. For when the asking surfaces will be available and on which plans, the current position is <availability - founder to fill>.