A source can answer a question without describing you
Suppose you consult a statute while preparing a document about your own situation. The statute can support a statement about what the statute says. It cannot support a statement that a fact about you is true.
Those are different propositions, even when they appear beside each other in the same workflow.
The distinction needs more than careful wording. It needs a boundary between two kinds of material:
- evidence brought in about the user or their work
- reference material consulted for rules, definitions, or external context
We keep the second group in named reference collections. A collection can hold legislation, product terms, or policy documents. Its passages can be retrieved and cited. They remain statements made by those documents.
Citation does not change ownership
A citation answers “where did this statement come from?” It does not answer “whose fact is this?”
| The source says | It can support | It cannot support |
|---|---|---|
| An application must contain a particular field | What the rule requires | That your application contained it |
| A policy sets a retention period | What the policy states | That your records were kept that long |
| A product has an eligibility threshold | What the threshold is | That you meet it |
If a statute says that an application must contain a particular field, that passage can support a sentence about the statute’s requirement. It does not establish that an applicant submitted the field. To establish that, the document needs evidence about the application.
The same boundary applies to terms and policies. A policy may define a retention period. A user’s own records would be needed to state when their data was retained. The policy is authority about the policy. It is not a record of the event.
This is why the answer model keeps case facts apart from passages. A person can provide a fact in the question or brief. That fact may be relevant to the answer. It is not thereby turned into something the reference material said.
The separation happens during ingestion
The wall is structural. Reference documents branch away from evidence extraction during ingestion. They are not admitted into the user’s evidence set and filtered out later by a label.
That timing matters. Once two sources share the same evidence pool, later checks have to recover their origin from metadata. Every retrieval result, citation, and generated sentence becomes a chance to apply the filter correctly.
With separate paths, the system has a simpler invariant:
- user evidence enters the evidence path
- reference documents enter the reference path
- reference passages can be retrieved and quoted in an answer
- reference passages cannot be minted as evidence about the user
The boundary is present before retrieval and drafting. It does not depend on a final permission flag.
Why a flag is too weak
A permission flag is easy to read as ordinary metadata. It can be copied, omitted, or applied only at the final step. A future query may select all chunks from a workspace. A draft builder may receive a mixed allow-list. A migration may preserve text but lose the flag.
These are ordinary failure modes in a system with shared tables and late filtering. The problem is not that a developer forgot a rule in prose. The problem is that the data model allowed the two meanings to travel together.
Structural separation narrows what later code is able to do at all. Reference material stays inside the collection it was admitted to. An answer can quote it as a passage from that collection. The material that supports statements about a person is reached by a different route entirely.
This does not remove every possible bug. It gives the bug fewer places to hide and gives validation something concrete to check.
What the reference answer is allowed to say
What a reference answer may assert is deliberately narrow. A finding must cite a retrieved passage. Its quote must appear in that passage. Numbers, amounts, dates, durations, and proper nouns must be licensed by the cited material, subject to the defined question and case-fact rules.
The answer can therefore say what the retrieved document states, supports, opposes, or defines. It can join clauses with ordinary reasoning. It cannot quietly add a particular that appears nowhere in the cited passage.
There is also a useful piece of metadata discipline. A source’s “as at” date and URL describe the document. They are held apart from the passage text. A display date cannot become a date that the clause itself asserts.
That separation prevents citation metadata from swelling the pool of facts a passage appears to license.
The claim boundary is the point
The goal is not to make reference material unavailable. The goal is to keep its role clear.
A statute can be cited as a statute. A policy can be cited as a policy. A product term can be cited as a term. A claim about the user still needs evidence about the user.
The safest boundary is the one that exists before a model retrieves, combines, or drafts. Once reference and user evidence have different paths, the resulting document can show both without confusing what each one proves.