Unit 07.00: Where a grounded system still invents things
Retrieval bounds what the model *saw*. It does not bound what the model *writes*, and the gap between those two is where grounded systems still hallucinate.
Extrapolation and merging
Two failure shapes account for most of it. Extrapolation adds a plausible detail the context never mentions. Merging combines two nearby concepts into one claim, so a policy about refunds quietly becomes a policy about refunds and exchanges.
The table below compares three candidate answers against the context, listing the substantive words each one introduces.
CONTEXT
[c1] Refunds are allowed within 7 days of purchase.
ANSWER WORDS NOT IN THE CONTEXT
"Refunds are allowed within 7 days." none
"Refunds are allowed within 7 days, and are
processed instantly." instantly
"Refunds and exchanges are allowed within 7 days." exchanges
All three cite c1. All three are on topic. "Processed instantly" and
"exchanges" are both absent from the context and both appear in a
confidently cited answer.
The grounded answer introduces nothing. The extrapolated one introduces instantly - a processing-time claim absent from a context that says only when refunds are *allowed*. The merged one introduces exchanges, a different policy the context never mentions.
All three carry [c1]. All three are on topic. The citation is real in every case, and the only distinguishing signal is which words appear in the answer but not in the source.
The mistake this prevents
The mistake is assuming a citation constrains content. It constrains nothing - attaching an id is a formatting operation. Two units from now this same word-level comparison becomes a check you can run on every answer automatically, and it exists precisely because the citation itself guarantees so little.
Takeaway
Grounding limits the input, not the output. Extrapolation and merging both produce answers that cite correctly and assert more than the evidence supports, which is why the comparison has to run at the word level.
