Writing
Mostly about things that looked fine and were not.
Where a number comes fromevery piece here
A run I didMeasuredQuoted
A vendor's figureQuoted
6 pieces
Reading a spec while it is still being written
A drawn sprite takes four minutes to arrive. Filling that wait with the sprite itself rather than a progress bar turned the order of two keys in the format into something load-bearing, and moving one of them puts the first pixel at 100% of the document instead of 46%.
A job that runs on time and does nothing
Building a dead man's switch for unattended jobs, and the two bugs that nearly shipped in it. Both were the monitor inventing certainty it did not have, in opposite directions.
What the judge does instead of judging
I scored eighteen model answers by hand and compared myself to an LLM judge on the same rubric. It agreed with me 94% of the time on the checks and 39% on the one criterion that needed judgement, where it reached for answer length instead.
When the model keeps asking for the same thing
A tenth of my draws were failing. The model was not choosing badly, it was asking for something the format could not express, and rewriting the prompt made it worse.
Testing a Godot game without opening it
Sixty-one headless suites and 1,740 checks on a colony sim, and the four bugs they found. None of the four produced a visible failure, so no amount of playing would have caught them.
A per-token price is not a price
Four models, four equally good answers, fifty-one times the price, and the field that would have under-reported one of them by a hundredfold.