Writing from the founder of synthetic cfo on synthetic financial data, forensic accounting and how to evaluate a generator.
A few days ago I wrote that OpenAI had conceded it has no standard for when to disclose misalignment. That same day it published one. It builds two of the four things a disclosure regime needs, and the two it cannot build for itself are the two that do the work.

A researcher's resignation set off a week in which the frontier labs asked for embedded evaluators, tests a model cannot see coming, and a licensed profession to run them. Audit has done all of it for a century, and the hard lessons are the ones being skipped.

Labelled occupational fraud data is scarce, because identifying fraud inside real enterprise ledgers requires both known cases and expensive expert annotation.

