AI, data and decisions

The criterion

Whoever decides must be able to get back to the data. Not "the model is explainable" in the abstract, but concretely: this number came from these rows, transformed this way, with this version of the model, on this date. If that chain cannot be reconstructed, the system is not ready, however good it is.

The work, deliberately ordinary

Declared data provenance, written transformations, checks that are able to fail: a check that cannot return a negative result is not a check, and a record of what the system answered and when.

On the AI side the question is not which model but which decision. A system that suggests and a system that decides need different guarantees, and the difference has to be written down before the tools are chosen.

How it can be checked

A decision system is checked by making it do the sum again: whoever reads a number must be able to trace it back to the query, the data and the version of the model that produced it, and get the same answer tomorrow. What we hand over carries that traceability with it, rather than an appendix promising it.