AI trading bot gave different answers to the same chart — temperature was never set
A developer built an AI trading copilot and posted it in another community for feedback. One commenter said: if you upload the same twice and get two different support levels, that's a dealbreaker. Checking revealed exactly that problem.
The vision call that reads the chart had no set at all, so it ran at the model's default, which turned out to be maximum randomness. The second call, which writes the final analysis, was running at 0.4. As a result, feeding the same chart into the pipeline on different runs could genuinely produce different support levels.
The developer had tested the product for months without ever noticing, because self-testing naturally involves grabbing a fresh chart each time rather than re-uploading the same one. The fix took two lines: setting to 0 and adding a fixed seed on both calls, shipped the same afternoon.
Key points
- Leaving unset makes the API fall back to a default that's often high-randomness
- Testing with the same input repeatedly is the only way to catch this kind of yourself
- For analysis/judgment calls, consider 0 plus a fixed seed as the default
- A single line of outside feedback surfaced a bug that months of self-testing missed