We ran the engine on real data. It refused to invent findings.
We put the frozen decision engine through a real, publicly published gambling dataset — 138,928 payment transactions. The dataset is payments only: deposits and withdrawals, no bets, decisions, or risk signals. So the engine could not evidence a single risk finding — and it produced none. Zero findings is the correct answer, and the point: it declines to fabricate what the data cannot support.
This is a discipline check, not a detection benchmark — and not a certification or a passed audit. A payments-only file cannot demonstrate fraud, capital, or responsible-gambling detection; validation on bet-level operator data remains part of a pilot.
Public, attributable, checksummed
Ghaharian, Ficco, Kim & Bernhard (2023), an online-gambling transaction dataset published on Mendeley Data (9j5gcygnwg) under CC BY 4.0. Used with attribution; the exact file we ran is fixed by SHA-256 so the run is reproducible against those bytes.
ad5e9aaf4ccfaec247549cef5c20a294db746b5dceeee054ca59b13bb3720bcb
Observed shape
- Transactions
- 138,928
- Distinct players
- 2,361
- Period
- 2019-03-01 → 2020-03-02
- Columns
- 5 (payments only)
Result on the frozen engine
- Engine (frozen)
- Ledgerproof risk engine v0.1.0 — unchanged
- Outcome
- Compatibility Report
- Findings produced
- 0
- Runtime errors
- none
The engine was not modified to run this dataset. On a payments-only file every analysis module reports UNAVAILABLE, which yields a Compatibility Report rather than any finding.
Payment flow (context, not a finding)
- Deposits
- 8,289,364.71
- Withdrawals
- 6,021,914.57
- Net flow
- +2,267,450.14
Amounts are in the file’s own units — the source carries no currency field, so no currency is claimed. Deposit and withdrawal counts sum to 138,928, matching the row total exactly. These are descriptive totals, not a Ledgerproof output.
Two places our observed data differs from the paper.
The file we ran does not match the companion paper’s summary in two respects. The file is the source of truth; we report the difference rather than quietly adopting the paper’s numbers. Neither affects engine behaviour.
| Field | Observed (file) | As published (paper) |
|---|---|---|
| Distinct players | 2,361 | 2,286 |
| Period end | 2020-03-02 | 2020-02-29 |
What this shows
- The engine runs on a real public dataset without modification.
- Evidence discipline holds: no required data ⇒ no finding, no fabricated number.
- It fails safe — a Compatibility Report, not a crash and not a false positive.
- The run is reproducible: fixed file (SHA-256) → recorded summary and log.
What it does not show
- It is not a certification and not a passed audit.
- Payments-only data cannot demonstrate fraud, capital, or player-protection detection.
- Detection quality is measured on bet-level operator data during a pilot — on your figures, not ours.
- No commercial, regulatory, or ROI claim is made from this dataset.
Reproduce it yourself → the full artefacts (summary, run log, rendered report, checksum, mapping) and step-by-step reproduction are published as a public repository under CC BY 4.0.
Dataset: Ghaharian et al. (2023), Mendeley Data 9j5gcygnwg, CC BY 4.0. Figures are OBSERVED from the run’s recorded artefacts and cross-checked against the file’s SHA-256. See also Explainability & Audit for how a real decision is evidenced.