Phil Pierce · Defensible AI decisions
I build AI systems that can defend their decisions.
To an auditor, a regulator, or a court. Inventor on pending patents in AI belief-memory, and twenty years running security decisions inside two of the most heavily regulated companies in America.
Eleven facts. Each one true at some point on the horizontal axis, each one learned at some point on the vertical. The verdict is not a property of the evidence — it is a property of when you asked.
Tools
The spec →Four minutes
DDR-1 conformance assessment
Fifteen questions about one of your AI systems. Returns a conformance level and the specific requirements you fail.
Run it →Bring a CSV
Calibration auditor
Predictions and outcomes in, reliability curve and calibration error out. Runs in your browser — nothing is uploaded.
Open it →Writing
All essays →August 2026
6 min
What Did You Know, and When Did You Know It?
Every inquiry into a bad decision converges on the same two questions. Most AI systems are architecturally incapable of answering either one.
Read →August 2026
5 min
A Confidence Score Is Not a Reason
0.87 is a summary of an argument. Under scrutiny, the summary does not survive and the argument was never written down.
Read →August 2026
6 min
You Can Buy Detection. You Cannot Buy Judgment.
Vendors are structurally incapable of selling you the decision layer, because the input it requires never leaves your building.
Read →The list
New essays, occasionally.
Evidence, memory, and machine decisions that have to hold up. No schedule, no filler.
Contact
If you are working on this problem, I would like to hear about it.
These are the questions I spend my time on. If you are building or buying a system that makes consequential calls and cannot explain them, write to me.
phil@philpierce.com