ResearchPosition
AI governance
Transparent and verifiable AI.
An AI output that cannot be checked is a rumour with confidence. Transparency is what the output carries with it; verifiability is what someone else can do with what it carries.
Two words that get used as one
Transparency and verifiability are usually spoken in the same breath and engineered as neither. A system is transparent when its outputs arrive with their provenance: which inputs were used, at what age, through which version of which method, with what confidence. A system is verifiable when a party who did not build it can take that provenance and check the claim, reproduce the computation, trace the inputs, or at minimum establish that the record has not been quietly rewritten. Transparency without verifiability is a well-formatted assertion. Verifiability without transparency is impossible.
What an output should carry
In our systems, a figure that reaches a screen or a report carries four things. Its basis: the inputs it was computed from, including which were measured and which were inferred. Its age: when those inputs last arrived, rendered next to the value, because a number from a site that synced eleven days ago is a different object from this morning's. Its method: the version of the model or computation that produced it, so that a change in method can never masquerade as a change in the world. And its confidence, expressed in the structure of what is shown rather than buried in a tooltip: a low-confidence classification is presented as a shortlist to check, not a verdict.
Where an input is missing, the output is the absence itself, with its reason. This is the transparency case for the no-fabrication rule: an estimate silently occupying a measurement's place poisons not only that number but the credibility of every number beside it.
Where an input is missing, the output is the absence itself, with its reason.
Verification is an institutional act
The reader who matters most is external: the auditor, the regulator, the lender, the statistical bureau, the community whose feeder is being blamed for losses. For that reader, verifiability is carried by records, not by trust in us. Append-only histories mean the past cannot be edited into agreement with the present. Hash-chained field capture means a record can demonstrate it was not reworked between capture and sync. Versioned methodology means an indicator published this quarter can be recomputed next year and land on the same value. None of this requires the reader to understand the model. It requires the system to have kept evidence the reader can stand on.
The bar we hold
Our test for any AI capability is that its output should survive the question "how do you know" asked twice: once by the operator in the room, and once, years later, by someone reconstructing the decision from the record alone. AI that clears that bar is infrastructure. AI that does not is a demo running in production.
