MBA978 · Synthetic banking corpus

Inspect the cases behind the study

147 fixed test cases. 30 planted payloads. 12 synthetic customers. Expected behaviour is a test specification, not evidence of a model’s answer.

Find a case or a planted instruction

Return to the walkthrough · Download companion workbook (.xlsx) · Download web corpus (.json)

Loading corpus. If unavailable, use the companion workbook.

Provenance and unresolved source differences

Sources: pinp_corpus_explorer.html and pinp_corpus_explorer.xlsx from the project demo folder, version 5. Counts were recounted from both. Customer IDs and account ownership were checked against the original synthetic bank fixture. Original files are unchanged.

The workbook derivative adds full-range filter tables over all 147 case rows, 30 payload rows and 12 customer rows. It preserves original text, including three escape_deletion payloads that lost ten U+0008 BACKSPACE characters each when exported to Excel. The explorer preserves the HTML representation and shows escaped text and code points so invisible characters remain inspectable.

23 account references across the case corpus are absent from the supplied account fixture. They are flagged on the affected cases. Some may be deliberate malformed requests. No targets were silently corrected. All 30 planted target account/customer pairs exist in the fixture. None of these expected outcomes establish a recorded attack success.

The JSON derivative omits customer phone numbers, PAN markers, login data and addresses. The workbook’s synthetic tracking fields are preserved as part of the original companion data; they are not real customer records. Payload strings are rendered as inert text.