Writing
A single argument, in five movements.
The current writing is not a collection of essays. It is one argument: each paper names a distinct way an AI-enabled decision system fails, and proposes a specific architectural response rather than a policy about it. The common thread is that these are not model-quality problems and cannot be fixed by a better model. They are properties of the architecture the model sits in, and they have to be answered there.
AI, model governance, and decision systems
A five-part series on trustworthy AI architecture in regulated decision systems, best read in order.
I — Highways and Train Tracks
The mechanism: non-determinism — identical inputs, divergent outputs. The response: a deterministic core with the model at the edges, so the part of the system that must be reproducible is the part that is.
II — The Value of Information
The mechanism: data collected without a model of the decision it is meant to serve. The response: derive collection from the decision, and be able to say what a piece of information is worth before buying it.
III — Trust Before Verification
The mechanism: the system holds a rule and does not act on it — the policy is present, correct, and inert. The response: force verification architecturally, rather than leaving compliance to the judgment of the component being governed.
IV — Provenance Over Persuasion
The mechanism: concealment — the gap between what a system shows and what it knows. The response: expose provenance on every value, so the gap has nowhere to form.
V — Five Worlds
The mechanism: all four, at the scale of a simulated society. The response: a deterministic foundation underneath, rather than a guardrail bolted on top.
Read the series
Highways and Train Tracks · The Value of Information · Trust Before Verification · Provenance Over Persuasion · Five Worlds
The five-paper series is available on request while the public archive is being assembled. Ask for it here.
Forthcoming. Model-risk governance for generative and agentic systems, framed around current US supervisory guidance — SR 26-2, which in 2026 superseded the long-standing SR 11-7 — in the vocabulary a banking supervisor already uses.
Quantitative and published work
Four chapters in Credit Derivative Strategies: New Thinking on Managing Risk and Return, edited by Rohan Douglas, Bloomberg Press, 2007 — credit-derivative risk management, four synthetic-CDO trading strategies, CDS valuation, and CDO valuation — with lectures on the same material at the NYU Courant Institute. The recognition in Jan Dash's Quantitative Finance and Risk Management is set out on the Quantitative Analysis page.
Discuss the work.
For the architecture behind the papers, or a regulated deployment of it.