Notes on building AI for decisions that matter.

A working paper behind OENRA — why a wine score answers the wrong question, how to put a wine’s chemistry and a person’s palate on the same eight axes, and where an honest model has to admit it can’t see. No results are claimed; the method is grounded in the literature and stated with its limits.

10 min read

Accuracy is one number, and it hides the thing that actually decides whether you can trust a model in a high-stakes loop: does it know when it’s right?

4 min read

A detector is never a finished object: the day it ships, someone starts routing around it. Notes on treating a detector's decay as a half-life, and what that means for high-stakes AI that has to assume an adversary.

4 min read

The most underrated capability in high-stakes AI is the ability to say "I don't know — this needs a human." A system that always answers is easy to build and hard to trust.

2 min read

From nine to thirteen I reverse-engineered games and ran a cheat operation against professional anti-cheat systems, until Riot's lawyers ended it. The habit of taking adversarial systems apart is now most of how I think about building AI.

4 min read

What actually works when you decompose a build across a team of AI agents — parallel research, bounded implementation, and adversarial review — and what still has to be yours.

3 min read

A go/no-go call is a bet on the worst plausible outcome, not the average day. Risk matrices score the average — here is what expected-shortfall thinking borrows from quantitative finance to fix that.

5 min read

In low-stakes AI a wrong answer costs a retry. In legal and aviation work the cost is asymmetric and delayed — so the output has to carry the rule, the fact, and the source that produced it.

4 min read

Most "AI for X" products are a chat box pointed at X's documents. But professional work is not answering a question — it is completing a repeatable unit of work with a decision and a paper trail.

4 min read