Jean-Luc Martel
Why an older discipline's apparatus survives contact with quantitative work: financial services technology, build-operate-transfer, and autonomous systems.
I keep this corpus because the most useful criticism of a quantitative result usually arrives from a discipline with no interest in markets. Textual critics, genealogists, chronologists and archaeologists spent centuries building tests for corrupted lineages, manufactured provenance and survivorship, under review considerably more hostile than a model validation committee. Their content does not transfer. Their procedure does.
My working life is technology for financial services — insurers, pension funds, wealth managers — much of it build-operate-transfer: stand a capability up, operate it, hand it to the people who will keep it. Handover is a merciless reviewer of analysis. A result that holds only while its author is in the room was never a result, and watching that happen often enough is what made me interested in tests that can be specified before anybody has seen the answer.
That is the whole apparatus here. Name the failure class. Name the near-twin that would produce an identical-looking result. Then name the observation that separates them. An entry that stops after the second move is interesting and unusable.
My current work is AI-assisted development and the security and governance of autonomous systems, where the problem recurs in a new register: a system emitting confident output far faster than anyone can construct the null it ought to be judged against.