Audit-grade is not a marketing word. It is a posture defined by what an external auditor must be able to prove from your data alone, without your engineers narrating it for them.
The four properties.
- Append-only history. Every change to a financial entity is a new row, not a mutation. The reconstructed state at any past moment is queryable.
- Deterministic settlement. Run the settlement job twice on the same input and it produces the same output, byte for byte. Non-determinism in settlement is a defect, not a quirk.
- Signed-off change control. Every production change is associated with a person, a time, and an approval. "git log" is not enough; the approval is a separate, immutable record.
- Preserved evidence. The artefacts that made a decision (the input file, the configuration, the deployment manifest) are retained for the regulatory window, not the convenience window.
Where teams underestimate the cost.
The retrofit. Append-only is cheap to build into a new schema; expensive to retrofit onto a system that has been mutating rows in place for five years. The retrofit usually means: a parallel event-store, a back-fill of historical state derived from the current snapshot (which loses fidelity), and a year of running both systems while the team builds confidence in the new one. We have done this twice. It cost us ~$300k of engineering time each time, and that was a clean retrofit.
What the regulator actually asks for.
In our experience: a randomly-chosen transaction from three months ago, with the full ledger trail, the settlement file it appeared in, the configuration in force at the time, and the engineer who deployed the code that produced it. Producing this in five minutes is the bar. Producing it in five days is the difference between a routine audit and an enforcement action.
The negative test.
A platform that calls itself "audit-grade" and cannot answer the previous question in five minutes is not audit-grade. The phrase is doing rhetorical work the engineering has not earned.
Audit-grade is what the data proves on its own. If a human has to narrate, it is not audit-grade.