Running now · 1 Aug – 29 Oct 2026

The 90-Day Trial

The author is running the book's own framework on himself, in public. Every 30 days the results publish here — the baseline, the control, what it caught, and what it missed. The final edition of the book includes all of it.

If the framework doesn't work, you will read that here first.

Last updated 2026-07-31 · Results published Day 30, 60, 90 · Book ships November 2026
In plain English

What's happening here

The book tells you to spend 90 days tracking your own AI mistakes, build a way to catch them, and show your boss the result.

It seemed dishonest to ask that of you without having done it myself. So I'm doing it, August through October, on real work, posting the numbers here as they come in.

Three check-ins: day 30 (how often AI actually got things wrong for me), day 60 (what I built to catch it, and how it broke), day 90 (whether it was worth the time, and how the conversation with my boss went).

If it doesn't work, that gets posted too. Which is also why nobody gets charged until the book actually ships.

Why run it in public?

Part IV of the book asks the reader to log ninety days of AI-assisted work, build one control, measure it, and take the result to their manager. Asking that of a reader while never having done it is not a small gap — it is the book's most attackable position.

So it is being closed the only way it can be: by running it, on real work, with the numbers published whether or not they flatter the method.

A framework that has survived one honest execution is a different object from one that has only been reasoned about.

What gets published

DAY 30 · 30 AUGUST 2026
Baseline  PENDING

A logged personal error rate for AI-assisted work across 30 days. Outputs processed, outputs verified, errors found, and the single failure mode that fired most. This is the number almost nobody in any company has.

DAY 60 · 29 SEPTEMBER 2026
The control, measured  PENDING

One control designed, run, broken, and revised — with its time cost measured by timer rather than estimated. Including what it failed to catch and why the first version was wrong.

DAY 90 · 29 OCTOBER 2026
The observed multiple  PENDING

A Judgment Multiple built from observed inputs instead of modeled assumptions, with the full assumptions block visible. Plus the manager conversation, and what actually happened in it.

The rules this trial runs under

RuleWhy
Publish the missesA method that only reports wins is marketing. The Day-60 entry is explicitly about what broke.
Observed stays separate from modeledAnything measured gets OBSERVED. Anything derived stays MODELED. They never merge into one number.
Redact by defaultError classes, not employers. Nothing identifying any organization appears in the log.
No retroactive fillingNumbers reconstructed from memory stop being observed. Logged at the moment or not at all.
If it fails, it publishes anywayA null result is a finding. Preorders are refundable and unpaid until ship for exactly this reason.

Run it alongside

The Career Control Kit is free and contains the same three tools being used in this trial — the AI Error Log, the Pre-Ship Verification Checklist, and the 90-Day Tracker.

Start your own baseline this week and you'll have ninety days of your own data by the time the book ships. Compare notes in November.

Preorder

Reserve a copy and the Day-30, Day-60 and Day-90 results come to you as they publish. Founding price $29, no payment until ship.

Reserve My Copy