For institutions
Auditable AI decisions
When AI moves near order flow, the question a supervisor asks is not whether the model is clever. It is what the model was looking at, and whether it could have known that at the time.
Where we stop, stated first
TWMD has no order path. It holds no broker credentials, sees no positions, sizes nothing, and cannot place, amend or cancel a trade. Anyone selling you an auditable-trading product is selling you something with a much larger surface than this, and the difference is worth establishing before the rest of the conversation rather than after.
What can be evidenced
Which data a decision saw, and whether it was knowable then. Every query can carry an as-of date, and the response states which knowledge-time field it filtered on. Every value resolves to a signed snapshot that a third party — including your regulator — can verify with a keyless endpoint and a script that imports nothing of ours. That is the data side of the audit trail, and it is the side that is normally missing.
Why supervisors are asking now
The FSB's consultation accepts that per-decision human review of agents is no longer workable and points toward AI monitoring AI. That raises an obvious question nobody enjoys: on what basis does the monitoring system trust the monitored system's data? A signed snapshot and an independent verifier are an answer to that question rather than a restatement of it.
What a pilot looks like
Take one strategy that already runs. Re-run it with as-of filtering and compare the results — if they differ, the original was reading the future, and that is worth knowing before a supervisor finds it. Then take a sample of its decisions and verify the underlying rows against the published snapshots. Neither step requires us to touch your order flow, and both produce something you can hand to a reviewer.
What we will not claim
A proof establishes integrity and origin, not correctness: if an exchange published a wrong figure, the proof attests faithfully to the wrong figure. Nothing here evidences that a model reasoned well, that a trade was suitable, or that an outcome was good. Those are questions about your model and your process, and a data vendor claiming to answer them should be treated with suspicion.
Read next
Not investment advice.