China Document AI: Preserve Identifiers Through Export
Leading zeroes and final digits can disappear after correct extraction. Follow six synthetic identifiers through export, import and a final round trip.

An AI document tool may recognize an identifier correctly and still change it when exporting the result. During a China enterprise-AI evaluation, follow a small set of synthetic identifiers from the source page through the saved record and into the receiving application. Compare the complete character strings at each boundary.
Define identifiers as labels
Use fabricated records, with no customer or payment information. For this fixture, every character is significant: leading zeroes belong to the identifier, and different strings identify different records. Agree that rule with the host before the demonstration.
Prepare three pairs: 00123 and 123; 000000 and 0; and 123456789012345678 and 123456789012345679. Give each record a separate test-row label so that the comparison does not depend on the identifier being tested. The six expected identifiers are distinct.
These values are labels, not quantities. Adding them, averaging them or converting them into a more readable number would not answer the business question. The long pair differs only in its final character, which makes a change near the end visible in an exact comparison.
Inspect three saved versions
Ask for the raw extracted value, the exported file and the value saved by the receiving application. Keep the original test-row label beside each. A screen that displays the right identifier does not establish that the downloaded file or destination record preserves it.
Compare strings directly. Check character count, leading zeroes and every digit. An exported value of 123 fails for the 00123 source even if both convert to the same number. Likewise, the two long identifiers must remain distinguishable after import.
Record where a change first appears. Correct extraction followed by a changed export is a different defect from incorrect recognition on the source page. The team responsible for the receiving application may need to investigate its import configuration even when the document model returned the right text.
Use the actual import path
Open the export through the workflow the customer intends to use. A CSV file, for example, does not itself guarantee that the receiving software will preserve every field as text. Inspect the imported value as well as the file contents. Do not treat a formatting change on the screen as proof that the stored value has been repaired.
Where the receiver requires a text column or an explicit field schema, have the host show that setting and repeat the import. Retain the chosen settings with the evidence. Avoid relying on an evaluator's manually corrected file: the repeatable workflow is what the customer will use.
Check distinct records after the round trip
Export the destination records again and compare them with the six references. Confirm that none has been merged into its paired record. A failed lookup, duplicate-key warning or silent replacement should be retained as a failed test outcome, not removed from the sample.
The document extraction guide examines ambiguous inputs and review. This fixture starts with unambiguous reference strings and tests preservation across system boundaries. The vision result identity guide addresses a different problem: matching asynchronous results to the correct inspection attempt.
NIST AI RMF 1.0 (2023), MEASURE 2.1 calls for documenting test sets, metrics and evaluation tools. Preserve the source fixture, each intermediate file, the import configuration and the final string comparison so the team can reproduce where an identifier changed.