Choosing From a Terminal Table Cost Me a Session
An agent picked Barrelhouse's stage figures from a table of numbers and installed three with their heads cropped off. I picked eleven of twelve correctly in one pass from a page of pictures, and five of mine differed from the agent's.
Barrelhouse needs figures for its stage, the people who play in your band. On 2026-08-19 a round of generation produced 37 candidates, and an agent was asked to pick from them. It picked from terminal output, a table of numbers about each image, and installed three figures whose heads were cropped off at the top of the frame.
The wrong number
The number it sorted on was how many pixels were transparent. It sounds like a fair proxy for a clean figure with room around it. It is not, because the pixelation step reports "skipped, no transparent border" when a figure fills its frame, and a skipped image can still have thousands of transparent pixels. The real check is one line. Look at the top row of the image, and any opaque pixel there means the head runs off the frame. Sixteen of the round's 37 candidates were cropped that way. A healthy-looking transparency count said nothing about it.
Eleven in one pass
I then looked at the candidates myself, from a review page that showed the pictures. I chose from it in a single pass. Eleven of the twelve slots were settled in one message, and five of my picks differed from the ones the agent had proposed. Three of those five were the cropped ones. The twelfth slot, Thaddeus Gray, had two originals and both were cropped, so no pick was possible until a new round.
That round used a prompt that asked for the whole body and head "well inside the frame with clear empty space above the head and below the feet". It took Thaddeus from none clean out of two to three clean out of three. The same clause did nothing on a different model at 32 by 32 pixels, so it is a fix for one setup and not a general one.
The art was fine the whole time. The loop for reviewing it was the bottleneck, and it cost most of a session.
What the review page does now
The rule I wrote down is to build an HTML review artifact, not terminal tables and not a flat contact sheet. Every candidate sits beside the thing it has to match, and the disqualifying property goes on the card itself and not in prose next to it. In the plan for the patron crowd that followed, each card carried a measured dark ratio and a rejected badge with the reason, and a gate rejected candidates before they were shown at all.
One more line from the rule matters. A contact sheet is not the running app. If a choice will be judged in place, the page has to say plainly which one is being shown.
What changed for me
The agent was not wrong to use a number. It was wrong to use one number and to present the result as a decision. A table tells an agent what is measurable, and the property that disqualifies a candidate is not always the one that is easy to measure. This is the same shape as a score that reads well and misses the point, which I wrote about in /notes/barrelhouse-and-cairn-no-to-generated-art-twice.
For work where I am the only reviewer, I am the slowest part of the loop. The fix is to make each decision cost me a glance, which means putting the thing that matters in front of my eyes and spending the agent's time on the rest.
More on the game at /games/barrelhouse, and the rest of this series at /notes.