Context
An identity check reads a German ID card through Azure Document Intelligence (prebuilt-idDocument). It kept offering the birth date as 08.12.1983 for a card printed 12.08.1983. That happened even after a fix that made the mapper parse the printed text day first. The tests for that fix passed, and the swap stayed.
Insight
A date-parsing fix can pass every test and change nothing. The provider swaps day and month inside its normalised valueDate, before your parser ever runs.
- Front only: classified as
idDocument.nationalIdentityCard, andvalueDateis correct. - Front and back stacked into one image: classified as
idDocument.passport. The machine-readable zone is lost, andvalueDatecomes back month first.
| Field | content |
valueDate |
|---|---|---|
| DateOfBirth | 12.08.1983 |
1983-12-08 |
The mapper read valueDate first and fell back to content only when it was missing, so the day-first fix was never reached. Running the captured response through the pre-fix mapper proved it.
Fix
Prefer content when it matches a dotted day-first format (dd.MM.yyyy, d.M.yyyy); otherwise keep valueDate. Don’t treat slashes as day first, because they’re ambiguous on US documents. They stay with the provider’s reading.
string[] printedDayFirst = ["dd.MM.yyyy", "d.M.yyyy"];if (DateOnly.TryParseExact(content?.Trim(), printedDayFirst, CultureInfo.InvariantCulture, DateTimeStyles.None, out date)){ return true; // printed text wins}// else: valueDate, then the other fallbacksInvestigation traps
- Capture what the reader actually sends, which is the composed image bytes, not the file the user picked.
- Write test fixtures in the shape the .NET SDK serialises, not the REST shape:
// SDK shape: what the mapper seesBinaryData.FromObjectAsJson(analyzeResult);// -> "ValueDate":"1983-12-08T00:00:00+00:00", "Content":"12.08.1983"
// REST shape (not what the mapper sees)// -> "valueDate":"1983-12-08", "content":"12.08.1983"