A regulator described the work. It didn't say who checks the AI.

In April, the Dutch Authority for the Financial Markets published a 30-page study of AI across the trading lifecycle. Its description of post-trade could be our own: “fragmented systems, manual reconciliations, and settlement breaks that cost firms time and money.” It names exceptions, “the discrepancies between trade records requiring human intervention,” as the bottleneck.

It is an exploratory study, and it says so: “It does not claim to be exhaustive or definitive.” It is also careful and worth reading.

But when it reaches the question of whether the AI gets it right, the answer is one conditional clause: “When properly validated, this supports faster processing, fewer errors, and reduced manual work.”

Validated by whom, against what, to what tolerance? The report does not say. The words assurance, inventory and model risk do not appear in it. Auditability, it concludes, “requires robust internal governance.”

A regulator described the work precisely. Who checks the AI doing it is left open.

“Validated” sounds like one property. It is at least three.

A preprint published in September offers a useful way to see it. Its author argues that independence in AI audit should be graded, not treated as yes or no, on three axes. Who controls the auditor. Whether the evidence can be checked by an outsider or is self-reported. And substrate: whether the auditor shares the system’s underlying model family, tools or guardrails, because if it does, it can fail the same way.

That third axis explains a control we keep seeing in operations: run two models, and escalate when they disagree. It is a sensible detection control. It is not an accuracy measurement. Two models from the same family can agree and both be wrong, and their agreement tells you nothing about it.

It is one author’s proposal, a first version without peer review, and not about finance. But the vocabulary is better than the one most of us use.

Two models agreeing is one witness, not two.

The test applies to us too.

Our early datasets were built with AI assistance, and the correct answers for AAL-D-001 were set by the author with that help, with no independent human reviewers. We already say so on the dataset page. By the substrate test, the next question is which model assisted, and whether it belongs to a family we also test. We will state that on each dataset page.

Our scoring is deterministic code, which no model can influence. But the answers it scores against are only as independent as the process that built them.

An evaluator that won’t grade itself on independence shouldn’t grade anyone else.

Take any AI validation your firm relies on and grade it on all three axes. Who controls the checker? Can an outsider verify the evidence, or is it self-reported? Does the checker share a model family, a toolchain or guardrails with the system it checks?

Score it by the weakest of the three. A strong answer on two does not rescue a fail on the third.

Sources: AFM, AI in Capital Markets: Balancing Innovation and Integrity (April 2026) · Ghanem, An Independence-Graded Audit Protocol for Agentic AI (arXiv, September 2026)