Writing
Essays on AI systems, healthcare decisions, and institutional design
I write about who absorbs the failure when an automated decision is wrong (the clinician, the claimant, the person still waiting in the queue) and what it takes for the correction to reach them.
The writing is part of the product work: I work out what software has to guarantee before a hospital or an agency will depend on it. The products on Current work test those arguments.
8 selected here, out of 230+ at The Crumple Zone, where each was published first. These are the ones that remain relevant beyond the week they were written.
Start here
The Official Record Is Late
A structure can be correct on paper and dangerous in practice — here is where that observation came from, and what I’m building now
If Every User Is a Potential Threat
People are not becoming dishonest. They are becoming game-theoretically optimal for the environment they have been placed in.
Toothless Ethics: Why Principles Don’t Stop Machines
A guide to the difference between moral language and structural constraint
More essays
Nobody Points "Next Best Action" at Themselves
Theory of Constraints, career-capital theory, RICE scoring, decision journals, and XP gamification, run as one script instead of five separate books.
Don’t Let Reassurance Do Engineering’s Job
Why “we care” substitutes for obligation — and how delay gets disguised as kindness.
You Don’t Have the Right
In this Age of Appeals, you have the paper right. And a stamina test.
How to Design for Cognitive Scarcity
Stop designing for the idealized “Hero User.” Build resilient interfaces that work when your user is stressed, tired, and operating on 15% battery.
Pending: The Political Economy of Waiting
The loading screen is the most powerful weapon in the modern state
How it works in practice
The essays argue the case. These pages describe the practice: what gets checked before an AI system is used, how its evidence is kept current, and what an institution needs in place to correct a mistake.
Framework
Deciding when an AI system may be used
What has to be written down before anyone relies on an AI system: the claim, the evidence behind it, who can challenge it, and what happens when it turns out to be wrong.
Operating model
Approval, correction, and recourse
The four stages from “not authorized” to “routine use”, the six-step review that runs when evidence changes, and the conditions that take a system back out of use.
Current work
Ambit, Refract, and Fast Harm
What an agent has been approved to do, whether its sources have changed since it read them, and whether a correction reaches everyone the error touched — alongside NextConsensus’s evidence work.
Recurring questions
- Why does an AI system punish faster than it can explain?
- Who is accountable when a model makes the decision?
- What infrastructure is missing between “AI can do this” and “AI should do this”?
- How do institutions maintain the ability to challenge automated decisions over time?
I keep writing about one pattern: people with less power absorbing the uncertainty, friction, and failure created by people with more control. Read the research program →
Revised