Trailing space deterministic, comfirmed A/B test

Howdy gents, who would have thought.

Model: gemini-3-flash-preview, thinking: low, temp: 0.

Last night I was testing a new OCR prompt, generating bboxes n such. I probably the exact same prompt 60 times on the same photo. Yes, yes, I was looking for outlier generations. In any case, today I was totally baffled at the results.

Three objects dropped out of 80% of scans, it was the 108-110 item out of 156 located near the bottom slightly right of center. This was completely unexpected, I had run this prompt 15-20 times through console with exactly the same setup and had not once seen that particular bimodal response.

Before doing a deep dive, I decided to check to see if the prompt I provide claude code was the same down to the byte… and there it was, a trailing space. Surely this couldn’t be it… but it was.

I just finished running running 20 more tests 1/20 had the predominate bimodal result. The bimodal result was achieved when sending requests every 5s. the next 15 tests, I introduced a 1min cool off… 0/15.

Anyway, if you’re reading this, you’re not crazy, you’re just a nerd.

Well check back occasionally for comments, comments doubtful :joy: