Use document retrieval when the question depends on language in a policy, contract, or note. Use a governed calculation when the question depends on a measured result the business has already defined. Use both when a reviewer needs the number and the passage, and keep them in one evidence packet so the narrative cannot treat either as the cause of the other.
The semantic-intelligence article explains why a term needs a governed meaning. This article is the routing decision that comes after that meaning exists: which evidence is allowed to answer which part of the question.
Synthetic energy example. The passage does not establish why the operational result occurred.
Two kinds of question
“What does the contract say about a standby rate?” is a question about a passage. A retrieval step can return a clause, a page, and a document version. The reviewer can read the words. The system should not turn that passage into a rate calculation unless a governed rule says how.
“Which drilling periods meet the approved exception rule?” is a question about a measure. The Energy blueprint keeps that measure in an approved rule and a review list. Retrieval of a nearby document does not replace the rule.
A hybrid question sounds like one sentence and contains both. “Does this period’s governed exception sit next to a contract clause about equipment?” needs the measure and the passage. The packet can hold both. The passage does not explain why the operational result occurred, and it does not authorize a contract action.
What retrieval is good at
Retrieval, including designs often called RAG, is a way to attach source text to an answer. It is useful when the evidence is a sentence, a definition in a policy, or a clause with an effective date. The limitations to design for are specific. A vector similarity score is not a measurement of business magnitude. A close paragraph can be about the wrong entity, the wrong date, or a superseded version. Similarity alone does not join a well to a contract, and it does not add a column of amounts.
That is not a claim that embeddings can never participate in a numerical workflow. A retrieved identifier might be an input to a later governed query. The limit is that the similarity step did not perform the calculation and should not be described as if it had.
What a governed calculation is good at
A governed calculation returns a result from an approved formula, grain, and scope. The Finance proof explorer is a local synthetic case of that pattern: budget, actual, variance, a contract version, and claim checks. It is not a warehouse query. It shows the shape of a result a narrative can cite.
The calculation does not know why the variance happened. A correct subtraction is not a cause. If the narrative needs a cause, it needs evidence a person can assess, or it needs to say the cause is not established.
A synthetic energy packet
Imagine a synthetic operating period, with no well name and no customer. The governed result says the period meets an approved exception rule because a required time code is missing. That result is an input to review. It is not a finding until a person confirms it. Separately, a retrieval step returns a short synthetic passage: “Standby time is reviewed against the approved field ticket.” The passage is not a private contract. It does not state a dollar amount, and it does not say the missing time code was caused by standby.
The evidence packet lists three things: the governed result, the passage with its document version, and an explicit unresolved item — the cause of the missing code is not in either source. Interpretation may say the period is on the review list and that a standby clause exists in the retrieved text. It may not say the clause produced the exception. A person decides whether any action is even in scope. The energy article keeps the same boundary for drilling, land, and operational health.
Scope, time, and contradiction
Hybrid evidence fails in ordinary ways. The measure is for September and the clause is from a prior year. The measure is for one department and the passage names another. One source says a status is open and another says it closed, with no shared as-of time. The packet should show the conflict. It should not average the two statements into a smoother answer.
Missing evidence is a state. If the passage cannot be retrieved, the measure can still be reviewed on its own, and the packet should say the document was not available. If the measure cannot be computed because an input is missing, the passage can still be read, and the packet should say the number was not produced.
What to evaluate
After a design is in use, evaluate it with questions you can actually score. Did the retrieved passage match the entity and the date the reviewer needed? Did the cited number match the governed result? Did the narrative add a cause that neither source supported? Did reviewers send the item back for missing evidence, and was that reason recorded? Those are evaluation questions. They are not results from a study, and this article does not invent adoption or accuracy rates.
Where each path fits
| Question | Evidence | Do not conclude |
|---|---|---|
| What did the text say? | Retrieved passage and version | That a metric changed |
| What was the approved result? | Governed calculation | Why it happened |
| Do we have both? | Packet with unresolved items kept visible | That one source authorizes action |
Discuss one question that mixes a document and a metric. The contact page is the place to start. The architecture shows the approved-tool sequence the calculation side should follow.
A routing habit you can review
Write the route down before the agent is allowed to choose. One practical habit is a short table owned by the decision, not by the model vendor. For each question family, name the passage source, the calculation contract, the fields that must match before the two may appear together, and the sentence patterns that are forbidden. “The clause caused the variance” is a forbidden pattern unless a person has recorded that conclusion. “The clause was retrieved and the variance was computed” is allowed, because it reports the packet.
Time matching deserves its own rule. A contract passage has an effective interval. A governed measure has a grain, such as a month or a well-day. If those intervals do not overlap, the packet stays split. Showing them on one screen without the mismatch is how a hybrid answer becomes misleading. The same discipline applies to entity matching. A lease identifier and a cost-center identifier are not the same key because a retrieval score was high.
Contradictions should be first-class. If the governed status is open and the retrieved note says closed, store both and mark the conflict unresolved. Do not ask the model to pick the more fluent one. The reviewer can request another document, reject the item, or accept one source for a stated purpose. That choice is a human decision and belongs in the record described in the audit-trail article.
Evaluation can stay small. Take ten synthetic questions you wrote yourselves. For each, score whether the passage entity matched, whether the number matched the contract output, and whether the narrative added a cause. Keep the score with the questions. Do not translate it into a market claim or a percentage of production answers. When a question needs only a document, do not force it through the calculation. When it needs only a measure, do not pad the answer with the nearest paragraph. The route is successful when it refuses the evidence that does not apply.
Write the refusal in the packet, not only in a log the reviewer never sees. “No passage was retrieved for this entity” and “no governed result was produced” are complete statements. They let the next person continue the review without replaying the question. A design that only stores successes will look healthier than it is, because the hard cases never appear in the workspace.