Induced from Eric Ma’s follow-up conversation with Abhishek Bachchan at the KENX AI in GxP East 2026 conference (2026-09-15-1531-abhishek-llm-acceptance-criteria-regulated-industries), 2026-09-15.
The method
Ingest every regulatory document you work with, then spend a long session prompting the model to write acceptance criteria grounded only in what you have done for your own applications. When it returns a basic statement, instruct it to generalize. Then check novelty: search whether anyone else has said it. His result held up on both axes, original and still valid against the corpus.
Why it matters for regulated writing
Facing the pile of documents a regulated environment demands, the hard part is starting, and templated statements unblock the start. Eric called it the most practical talk of the day for exactly that reason. Those statements still have to name a population and a threshold; this is how you draft the criteria inside Validating an AI system in a regulated industry means validating a measurement system, and they still need A golden test set must include unhappy paths or the novelty is only verbal. The caveat Abhishek volunteered alongside: his six-month MIT class on LLMs was already six months out of date by the end, so the practice needs continuous refreshing even as the method holds. That refresh cycle is the writing-side version of AI validation evidence has a shelf life.