The cash desk software was being replaced, and it had to be proven that the new system counts exactly as the old one did. We wrote an emulator for every gaming client — twelve of them — ran thousands of bets through them and reconciled the reports of both systems line by line. Any discrepancy failed the build, so "it seems to add up" became a verifiable fact.
For every gaming client we wrote a headless emulator of its network contract — twelve of them. The emulator behaves like a real client but without an interface and in any number of instances.
A run is thousands of bets, and every win is reconciled against the odds tables taken straight from the backend sources rather than from the documentation: documentation goes stale, code does not.
The reports of the old and the new system are reconciled line by line. Any discrepancy produces a non-zero exit code in CI, that is, a red build. "It seems to add up" does not pass acceptance.