A hospital discharge agent mixes two patients in one answer. The grounding folder holds scans with no encounter identifier, so a file cannot be tied to one visit. A teammate wants a new model. What should the architect correct first?
Select an answer to reveal the explanation.
Short Explanation
Two charts dropped in one unlabeled folder will get read as one patient. You clean up the scans and give each visit an encounter identifier before anyone trains a new model. A longer instruction that says to keep patients apart does not label the files.
Full Explanation
Grounding cleanliness fails when source files cannot be separated into the records they belong to. These scans share a folder and carry no encounter identifier, so the agent mixes two patients because the data itself is not tied to one visit. The first correction is to clean and structure that set, with an encounter identifier on each scan, before anyone adds a model. Training a custom model on the mixed folder asks the model to guess an identity the files do not state, and it will guess wrong in the same way on the next unlabeled scan. Lengthening the system prompt tells the agent to keep patients apart and leaves the files unlabeled, so the retrieved text can still contain two visits. Retiring the discharge agent until every ward has its own vision model swaps a data defect for a pile of models and still does not label the scans. A model can only repeat the structure it is given. Exam caveat: cleanliness here is about one file belonging to one encounter, not about how polished the prose of the answer sounds. Operational check: attach encounter identifiers in a test folder, ask a discharge question for one visit, and confirm the answer uses only that visit's scans.