Skip to content
How much to read

Knowledge base

A pipeline reads my repositories and project folders every hour and writes these pages. A second model, which cannot see the first one’s reasoning and has no tools to go looking for agreement, checks each one against the source it came from.

68 pages here, out of a vault that is larger. Everything else is graded private, and the export refuses to write it. Numbers arrive with the caveat that was recorded next to them, and rows of a table that are not publishable are stripped out of tables that are.

Last exported 2026-09-22. How it works is written up in the case study, including the parts of it that can still be wrong.

Start here

7 pages

The working rules everything else follows from, and what each project is. Read these first if you read nothing else.

Where the model is kept on a leash

9 pages

For anyone checking whether I understand evals and retrieval or only the vocabulary. Each page is a place I decided what the model is not allowed to do, and how that gets checked.

What I cut, and why

7 pages

For founders and hiring PMs. Things I removed or declined to build, with the reason written down when I made the call.

Changed after watching the user

10 pages

Redesigns that started with how students and a dentist I know actually used the thing, not with what I had planned.

Making the data trustworthy

9 pages

Every feature above rests on data someone could get wrong. These are the checks that compare it against its source instead of against a number someone remembered.

Privacy and access

10 pages

What is kept out of public view, who is allowed in, and how the privacy claims are tested against what is actually stored.

Running without me

6 pages

For technical readers. Scheduling, recovering from a stopped run, fixing concurrency in the right order, and putting a ceiling on what a run can spend.

Records

10 pages

Timelines, numbers with their caveats, and the questions still open. Reference rather than argument.