Publication audit
Scope
The dataset contains transformed counterparts of the same 118 underlying cases used in the reported evaluation of the Crypto Accounting Bench research paper. Each task is matched one-to-one to an authoritative source row, and identifiers run task_0001 to task_0118 with no gaps.
The paper reports 246 non-zero reference lines: 113 two-line entries and 5 four-line realized gain/loss entries. Twelve stored answers also retain exact-zero template lines for provenance; the official scorer ignores exact-zero lines, so these do not change the paper population or grading contract.
Result
PASS: paper-population, privacy, answer-integrity, Viewer-projection and checksum checks passed.
| Check | Result |
|---|---|
| 118 source rows and 118 gold answers | PASS |
| 118 one-to-one source-to-public task matches | PASS |
| paper composition invariants | PASS |
| 246 non-zero reference lines (113 two-line, 5 four-line) | PASS |
| every journal entry balances | PASS |
| every answer and deciding account exists in its supplied chart | PASS |
| every prompt contains the supplied answer accounts | PASS |
| every rubric sums to 1.0 | PASS |
| every realized gain/loss recomputes from transformed visible evidence | PASS |
| no original transaction hash, address, source identity, private path, email or credential marker | PASS |
| Hugging Face Viewer rows exactly reproduce task artifacts | PASS |
Residual risk
Automated checks do not prove anonymity. Chart structure, accounting semantics, relative magnitudes and temporal spacing are preserved because they are part of the benchmark. An insider holding a source ledger may still recognize structural patterns. No direct join key to a real party or transaction is included.
Publication authorization
Technical sanitization does not establish legal or contractual permission to redistribute the source records; that is a separate determination made by Entendre Finance.
That determination has been made. This transformed derivative is published publicly under CC BY-NC 4.0. The non-commercial restriction is deliberate: the dataset derives from production accounting records, and redistribution for commercial use is not granted.