Finding
Some tasks are easier to fix than others
I sent outputs that failed back to models, told them what was wrong and asked them to fix it. Here is each task before, fixed and not fixed.
The V60
In a previous run Haiku 5.5 drew the paper filter under the dripper. I sent that failed run to the models and asked them to fix it. All models were able to fix the issue.
The knitting chart
In a previous run DeepSeek 4.1 Flash drew a bean that knits 2.4 times as long as it is wide. I asked the models to fix it and Deepseek did but Haiku did not manage to fix it.
The Venn
In a previous run Kimi K3 put the word milkshake on the edge of a circle. The job is to move that one word clear of every line and keep it in its region. Same I sent it to the models. Kimi fixed it in one run but not in a second run.
The spoon
GLM 5.3 FlashX drew a spoon lying across the mouth of the cup. Deepseek managed to fix it. GLM failed to fix its own previous mistake.
GLM did manage to fix it in another run.
| Repair | Fixed | Unchanged | Board |
|---|---|---|---|
| V60 order | 9 of 9 | 9 of 9 | retired |
| Bean knits true | 7 of 9 | 9 of 9 | retired |
| Label off the edge | 7 of 9 | 9 of 9 | live |
| Spoon inside the cup | 4 of 9 | 9 of 9 | live |
So I retired two tasks because two models cleared every run.
The 35 runs cost $3.24.