AI governance is converging on a documentation model: an inventory of systems, a statement of policy, logs of activity, a record of authorization, and an ever-lengthening list of supporting artefacts. This paper argues that the model has a structural ceiling. Documentation records what was intended, written, or approved; a defensible account of an automated decision requires something documentation cannot supply — the ability to reconstruct what the system was actually bound to do at the moment it decided. Adding artefacts raises the cost and the surface of compliance without crossing that line. We trace the pattern from a popular four-record framing, through a representative multi-regime compliance matrix and its expansion toward forty distinct artefacts, to its logical endpoint, and argue for a shift from documentation-centric to derivation-centric governance: from asking whether the records exist to asking whether the decision environment can be reconstructed as it stood at execution. The distinction is the difference between what was written and what was binding — and, increasingly, between a record and an admissible one.
Narnaiezzsshaa Truong (Fri,) studied this question.