23 September 2026
The Judge Wrote "Disconnected" and Let It Through
Three AI models judged the images my social-media bot wanted to post. What their scores missed, what their notes caught, and the price of letting no bad image through.
Dargel Solutions · Essays
What broke, and what changed.
Under the name Adrian Verdan, the publishing imprint of Dargel Solutions, we publish reports on our own automation — each one built on a specific failure, with the date, the log, and the change that followed. Every report is here in full; most are also on Medium. Free to read, no sign-up.
Newest first.
13 September 2026
The Recheck of the Recheck of the Documentation
An acceptance apparatus for AI coding agents spent four days checking itself. What the gate cost, what it caught, the door I built past my own gate, and what I switched off — every number from my own records.
30 August 2026
Five AI Reviewers Preferred My Prototype. I Rejected It.
Five repeated AI review runs unanimously preferred the more active film scene. Four also filed a Major; one of those reports exposed how the pictures made unresolved tasks disappear into an archive and called it progress. Why a 5–0 can choose a direction but cannot grant approval.
30 August 2026
My Test Documents Had No Pictures in Them
Fifteen rules, 457 passing tests, 77 of 77 mutants killed — and not one test document my layout checker had ever seen carried a figure it could measure, while 40 of the 74 pages in the guides I sell do. A corpus written to make the tests pass measures the author, not the world.
17 August 2026
I Ran the Same Audit Nine Times. Four Missed the Hole.
Nine runs on three fixtures I wrote myself, same prompt, same morning — the only variable was the model. The cheaper model missed a deliberately planted missing entry in four of five runs; the expensive one caught it in four of four.
11 August 2026
Six Products, Two Sales, One Customer
Six products, five weeks, and a single buyer who bought twice. Auditing my own website counters turned up one that was wrong — and a second that only surfaced because somebody else read the draft. I end with one claim fewer than I started.
Also on Medium →
5 August 2026
My Second Opinion Was Dead for Four Days. At Least.
The second-opinion runner exited with code 7 and wrote no file, for twelve days, unnoticed. In a log, silence and agreement look the same. The report corrects the number in its own headline.
Also on Medium →
3 August 2026
My Tests Were Green Because I Wrote Them
An adapter read camelCase while the interface delivered snake_case. The tests stayed green because they asserted exactly what the code did. The report sets out the signs of a check that only measures its own assumption.
Also on Medium →
1 August 2026
Every Safety Rail Worked. Nothing Happened.
From 23 to 28 July our marketing bot published almost nothing: three zero days on the 23rd, 27th and 28th, with three, two and one post on the days in between. No crash, no error message — on those zero days a daily cap, a circuit breaker, an image quota and a release gate, each set correctly on its own, came to zero together. The report works out the combined capacity.
Also on Medium →
The profile with all reports: Adrian Verdan on Medium →
