Trailing space IMPACTS output confirmed, by me, in A/B testing
Not sure if anyone will care, but a trailing space (extra spacebar white space) at the end of a prompt wrecked at least 12 work hours 😂.
Model: Gemini-flash-3-preview
Thinking: low
Temp: 0
All the sudden a battle tested prompt was showing a consistent bimodality I hadn’t seen before. Ran maybe 60 total tests, same image, same prompt, same model.
Great on console. Predictable 8/10 times, consistent boundary boxes (for OCR). Then I ran into an issue as I was running agent tests through my pipeline.
Three objects dropped on the bottom of the picture. They were objects 108,9,10 of 129.
Tried everything to figure out what was happening. Then I asked Claude code, is the prompt you are using byte identical to “x” prompt.
Answer, na… there was a singular byte difference. A trailing space. Agent thought it would be nothing but I wanted to test it anyway.
First round of five tests 4/5 perfect one instance of dropped objects. The I fired 15 more with a 1 min cooldown between tests
Total 19/20 perfect.
With the trailing space I got bimodality, dropped objects, 36/40 times.
Anyway, yer a nerd for reading. And that’s ok bc a blank space making your product fail is wild af