Replayable or inadmissible

Run the experiment: ask exactly the same question, twice, of the same model. The two answers will differ — sometimes at the margin, sometimes on substance. This is not a manufacturing flaw: it is a design choice. Generative models draw each word at random from a distribution of probabilities, and a measure of chance is deliberately kept to make answers varied, natural, creative.

For conversation, this is a virtue. For decision, it is disqualifying.

Nothing is reproducible, therefore nothing is auditable.

Every discipline of trust rests on reproducibility. A scientific experiment that cannot be repeated is not a result, it is an anecdote. An accounting calculation that yielded two different totals depending on the hour would pass no audit. A laboratory test that cannot be reproduced certifies nothing. Wherever responsibility is engaged, the rule is the same: what cannot be replayed cannot be proven.

Yet each day we entrust more analyses, syntheses and recommendations to systems whose every output, by construction, cannot be replayed. When a regulator, a client or a judge asks « why did your system assert this? », answering « it would no longer put it the same way today » is not a defence. It is a confession.

The law has already grasped this: the European AI Act requires traceability of decisions for high-risk systems — logs, records, the capacity to reconstruct what happened. One cannot trace what does not reproduce. Determinism is therefore not an engineer's preference: it is the entry condition for AI into the processes that commit an organization.

A verdict worthy of the name must be replayable — same question, same sources, same result, today, in six months or in six years, whatever models have come and gone in between. Replayable, or inadmissible: there is no third category.