Four status updates enter an organisation’s approved knowledge over a fortnight. An integration is delayed, a release date moves, a supplier contract remains in legal review, and a dependent change is waiting for approval. Each update is accurate, and each has an owner. Retrieval can return every one of them with its source. The harder question is whether they share a dependency that changes what someone should do next, because that dependency has no record and no owner.
An enterprise brain gives people and software governed access to an organisation’s knowledge. It remembers what the organisation knows, assembles evidence, supports reasoning and lets authorised work proceed. A reference architecture for the enterprise brain gives two of its capabilities a short contract.
Recall remembers; observations understand.
Recall returns approved claims and provenance. An observation describes a pattern across several claims, with the evidence attached. It is a knowledge-derived view that carries its own identity and lifecycle, together with its own serving controls. Approved knowledge remains authoritative.
Treat the observation as a map
Alfred Korzybski separated the map from the territory. An observation is a map of a territory made of approved claims, and the map may be redrawn at any time without changing the ground it describes.
An observation records its member sources, grouping basis, evidence, generated interpretation, lifecycle state and generation context, so a reader can open the material the narration draws on. The observation appears beside recall results, and it can answer a request about patterns directly.
The information classes have distinct owners. The knowledge source of truth holds approved claims, and the knowledge-derived view holds interpretation together with serving state, kept in an observation store. Governed control state holds reviews and corrections, and release remains a separate governed decision over what reaches a consumer.
An observation can be rebuilt from approved knowledge and governed control state without changing what the organisation has approved as true. A solution may phase delivery against a stable logical contract.
Fix the structure before the model writes
This architecture separates the structure of an observation from its prose. Deterministic analysis creates signatures, groups, stable identity and lifecycle events from approved claims, and the same claims and rules produce the same structure.
Cluster membership is fixed before the model is invoked. The model receives evidence keys and relevant excerpts, then writes a short account of what the sources jointly record. The narration is asked to stay within the evidence and to identify a conclusion that matters to a consumer: a decision, constraint, dependency or change of state.
A summary that only repeats the link structure adds no understanding and should not become active.
Grouping sources by the references they share is bibliographic coupling, described by Kessler in 1963, and deterministic grouping followed by constrained narration is established practice in graph-based retrieval pipelines. The addition here is that the narration carries identity, lifecycle and serving control.
Repeatability makes the grouping inspectable and testable. Meaning still depends on the cited material and governed judgement: relationships absent from the available structure remain outside the result, and a threshold can exclude a connection a person would recognise.
The opening example resolves here. The four updates share references to one supplier contract, so the deterministic pass groups them, and the narration states the conclusion the set records: the release date depends on a contract still in legal review. The observation names the leverage point, and the release owner’s next action is the contract review, with the evidence one link away.
The brain’s prose is a governed write
When the system persists its synthesis, it becomes a contributor to its own derived views. The prose appears beside approved claims and may arrive through the same interface, so the architecture must stop generated text from acquiring the authority of its sources.
Every narrated observation carries evidence, authorship, generation context and lifecycle status. The contract places four checks before activation: evidence completeness, policy conformance, potential contradiction and required oversight. Maker-checker and provenance predate generative AI, and here they apply to text the system writes about its own knowledge.
A potential conflict moves the observation to held status. Its structure and review evidence remain available to the control plane. Consumer surfaces receive none of the prose, and held observations do not enter the observation index.
The reviewer receives the generated claim, conflicting evidence, cluster context and available verdicts in one decision record. The verdict and outcome are persisted together, keyed by the decision, so a request sent twice returns the recorded result and acts once.
Correction follows the same control model. A reviewer may pin replacement prose through a recorded decision or request fresh narration with a correction note. Pinned prose carries human authorship in place of the model’s generation stamp, so the record shows who wrote it and under which decision. Both routes record the decision that produced them and keep approved claims unchanged.
Lifecycle rules complete the contract. Active observations may serve and embed, and held or retired observations stay outside consumer retrieval. A membership change triggers regeneration, a new structural identity creates a new observation, and dissolution retires the old one. Governed control state preserves pinned corrections through ordinary regeneration.
Human claim admission governs entry into approved knowledge, and observation lifecycle governs derived interpretation. Consequential use carries a requirement of its own: an oversight verdict and a governed release decision.
Separate observation retrieval from source retrieval
The observation store contains pointers to approved claims and prose about them. Generated summaries use an observation index of their own, and source retrieval holds approved claims. Evidence links always resolve back to approved claims.
Separate indexes keep generated prose out of source retrieval, so later generation cannot cite it as evidence.
Without that separation the loop closes on itself: a generated summary becomes retrievable evidence, a later generation cites it, and the system accumulates interpretation that stands on earlier interpretation, every link still resolving to something that looks like a source. Index separation also removes held and retired observations from serving without changing approved knowledge. An answer may cite an observation as interpretation, and factual claims cite the evidence behind it. Prose reaches approved knowledge only when a person proposes it through claim admission.
How to judge the layer
Judging this layer needs evidence beyond retrieval quality. These questions test whether a design that persists synthesis is governed, and each has a failure behind it. A grouping that cannot be reproduced cannot be audited, and a correction that regeneration erases returns the same error to the next reader.
- Do the same approved claims and rules produce the same grouping and identity?
- Can every material claim resolve to the exact evidence that supports it?
- Do serving and embedding exclude potential conflicts?
- Does the design keep generated prose out of source retrieval and stop it citing itself?
- Does a recorded correction survive regeneration and a change of model or prompt?
- Does the layer find known relationships across representative questions and corpus shapes?
- Does an observation change a decision, priority, sequence or investigation path for a defined consumer?
What the organisation gains
Maker-checker, provenance, lifecycle state, idempotent decisions and separation between authoritative and derived stores all predate generative AI. The observations layer applies those controls to a system that writes interpretations about its own knowledge.
The result is understanding the organisation can defend: for any sentence the system produced, it can answer which approved claims stand behind it and who agreed it could be served. The reference architecture and its reusable artefacts are published under CC BY 4.0 in the Enterprise Brain Reference Architecture repository.