What this is
A model can invent a permitting timeline that sounds finished. It can also produce one where every fact traces to a named source with a time on it, and the sequence logic came from a rules file a person wrote. On the page those two timelines can look identical. They are not the same kind of claim.
Provenance is tagging which bits were fetched, which were calculated, and which were generated. The same voice can carry all three. The failure modes are different.
Generated narrative, retrieved records, and deterministic computation can be concatenated into one fluent answer.
Provenance is tagging which spans were sampled from the model, which were fetched, and which were produced by code or rules a human authored.
Easy to mix up
This is easy to mix up with official-sounding writing. Tone is not a source. A generated docket number is still a guess, even if it is formatted like a record.
It is also easy to mix up with a link at the bottom of a paragraph that does not support the sentence. Provenance is about the span, not the footnote.
Fluency and formatting are not evidence. A plausible identifier with no fetch is still a sample from the prior.
A URL appended to a generated paragraph is not provenance unless the span can be traced to that source. Reliability is not uniform across an answer.
From work
A policy memo needs a date: when the comment period closed. Retrieval can pull that date from the docket. A script can compute “thirty days after publication” from a rule. The model can also write a date that looks like the other two.
If the memo cannot say which of those three produced the date, the date cannot be treated as a record. One of those objects is a fetch. One is a calculation. One is a generated date.
Statutory dates and docket numbers should resolve to a primary source and a retrieval timestamp, or to a deterministic function over stated inputs.
Interfaces that collapse retrieved, computed, and generated spans into one bubble hide the only information that would let a reader discount the generated parts.
What to ask next time
For this sentence: was it retrieved, computed, or generated? If the interface cannot say, do not treat the sentence as a record.
Ask for span-level tags. Retrieval fails on index and permissions; computation fails on spec and inputs; generation fails by sampling. Those errors are not interchangeable.