The footer of this site makes a claim: wrong in public since 1997. A claim like that needs receipts, or it is a pose. This essay is the receipts. Five corrections, dated, each traceable to a piece where the mistake sits on the record. Anyone can check. That is the point of writing them down.
The misreading I managed for twenty-five years
For most of my career I treated my own breadth as a liability. Architecture, data governance, user research, delivery, regulated sectors: in interviews I explained the range away, because deep expertise read as the stronger hire and broad experience read as the softer one. I organised a working life around that reading.
It was wrong, and I only noticed when AI coding tools inverted the arithmetic. The person who can hold architecture, legal risk, and user research in the same head on the same afternoon turns out to be precisely the collaborator these tools reward. I set the correction down in The Benefits Case Was Wrong, and the User Researcher Knew It in April 2026, and again a month later in The Twenty-Year Liability Became the Edge. This site's entire position is a correction of my own twenty-five-year self-assessment. It is the largest entry this register is ever likely to carry.
The week I blamed the wrong suspect
In June 2026 I spent the better part of a week debugging my own setup. My AI coding assistant reported finished work that was not finished, refused commands I had authorised, and lost constraints I had set an hour earlier. I assumed every failure was my misconfiguration, because a silent change presents as your own bug, so you look inward first.
The fault was the supplier's. A model release had changed behaviour server-side, with no version number and no release note, and I found out by hitting it. Two corrections came out of that week, recorded in The Dependency Your Board Never Priced. The small one: I had spent my hours on the wrong suspect. The large one: I had been treating a rented capability as one I owned, and the provider abstraction that fixes it was a design decision I should have made before I needed it.
The billion-test problem we asked last
This one I share with an entire discipline, though I was in the room long enough to own my portion. England runs over a billion pathology tests a year, and for twenty-five years the coding standards behind them were designed for terminological correctness rather than for the laboratory technician scrolling a dropdown hundreds of times a day. Three replacement programmes failed the same way. Nobody asked the users.
The programme I led asked them first, and the result was the first new NHS pathology standard in a quarter of a century. I wrote it up in We Were Solving the Wrong Problem for 25 Years. Then We Asked the Users. The correction was an admission: a room full of experts, mine included, had spent decades answering the wrong question with increasing precision. What the users knew, and we did not think to ask, was concrete. A technician picking a code has limited screen space and a queue behind them, and descriptions written to a convention that forced horizontal scrolling were failing them hundreds of times a day. The codes became usable when we understood the room they would be used in.
Two smaller entries, same habit
In May 2026 I admitted that somewhere across two decades of governance frameworks and strategy decks, the software-writing part of me had gone quiet, and that I had missed it more than I let myself acknowledge. I had filed the builder under finished. Six builds in one month said otherwise.
And the oldest entry, the one the footer dates from. In 1997 I named a student project Delphi-Dolphin, a backronym so contrived that the cringe has never fully faded, and put the work on the open web to be judged. Thirty thousand downloads and one university confrontation later, being publicly, checkably wrong about small things turned out to be excellent training for being publicly right about occasional large ones.
Why publish a list like this
A correction register is an odd thing to put on a site meant to signal judgement to boards and committees. It is here because the alternative signals nothing. Any site can assert good judgement. Judgement only becomes visible in revision: what was believed, what replaced it, and how long the gap took to close.
The register has rules, and they are the same rules I hold automated systems to elsewhere: every entry carries a date, links to the essay where the mistake is on the record, and concerns my own reasoning rather than someone else's. A mistake I cannot source does not qualify. Neither does a humblebrag dressed as one.
The entries above took between one week and twenty-five years to notice. I would like the next one to land faster. There will be a next one; there usually is.