A fintech AI screenshot tool learned that temperature 0 still needs full-pipeline determinism
famelebg29 · reddit · 2026-07-22
A builder of a fintech AI that reads trading-chart screenshots found that the same image could produce different support levels on repeated runs.
What caused it
- The vision call had no temperature set, so it defaulted to 1.0.
- A later stage in the pipeline ran at 0.4, so even if one step was deterministic, the final output was still re-randomized.
- The bug was easy to miss because the team rarely tested the exact same input twice.
What fixed it
- Set temperature to 0 throughout the pipeline.
- Add a fixed seed on every call, with the caveat that OpenAI only treats it as best effort.
- Test identical inputs in CI.
Main lesson
Determinism is a trust feature, not just an accuracy feature. A consistent wrong answer can look more confident than it deserves, so the next layer is grounding: reconcile model claims against source-of-truth data. In this case, that meant checking recognized chart pixels against real OHLCV candles and failing closed when they disagree.
The author also shares a debugging trick: if the reconciler fails, rerun it against the 2–3 most likely intervals. If one matches, the interval detection was wrong; if none do, the model’s read was wrong.
More from coding & agent
- Tencent Hunyuan’s Hy3 ranks #25 in Agent Arena and #2 among open models in Frontend Code Arena — arena · 2026-07-22
- Bindu Reddy says OpenAI and Anthropic already rely heavily on vibe coding — bindureddy · 2026-07-22
- Karpathy says Codex and Claude Code need user control over subagent roles and model choice — emollick · 2026-07-22
- Writing too few CLAUDE.md files keeps forcing models to rebuild context — TheZachMueller · 2026-07-22
- Telemetry can’t prove an AI intrusion was fully autonomous — cyb3rops · 2026-07-22
- Helmholtz Munich and TUM open fully funded PhD spots in trustworthy AI for science — zeynepakata · 2026-07-22