A validation step in the Ledger Intake pipeline rejects any extract_invoice call that fails the schema, then re-sends the same document together with the validator error message appended to the request. Across one week this recovers 96 percent of failures in which invoice_total arrived as a string carrying a currency symbol, but only 4 percent of failures in which invoice_number is required and the page has been cropped so the header no longer appears. What best explains that split?
- AThe cropped cases fail because the retry re-sends the document without the validator message attached, leaving the model with no more information than it had on the first attempt.
- BThe cropped cases fail because repeated requests are served from a response cache, so an identical document keeps returning the identical invalid arguments no matter what feedback is appended.
- CCorrective feedback tells the model how to reshape a value it already holds, which repairs a badly formatted number, but no feedback can recover an identifier the cropped page does not contain. Correct
- DThe cropped cases fail because a cropped scan pushes the request past the model's attention budget, so raising the context limit on those requests restores the missing header field.
Why A is wrong: Tempting because a retry that drops its corrective feedback is a genuine implementation bug worth checking for. The described pipeline does attach the error message, and the currency symbol cases recover precisely because that feedback arrives.
Why B is wrong: Tempting because caching does make retries look futile and is a real cost optimisation in production pipelines. A cache would suppress recovery on both failure classes equally, which contradicts the 96 percent recovery on the formatting faults.
Why C is correct: Correct. The two failure classes differ in what the retry can supply. A formatting fault is fully recoverable because the information is present and only the representation is wrong, whereas a missing header removes the information itself, so further attempts can only produce a guess.
Why D is wrong: Tempting because attention dilution on long inputs is a real limitation and cropping sounds like a size problem. Cropping removes content rather than adding it, and no context setting can reinstate pixels that were never sent.