
Fast, bounded model judgements can improve routing, but only a versioned decision contract should turn confidence into consequence.
The argument
A probability or score is evidence under a specific model and question, not authority to approve, block or escalate an action.
What is new
OpenAI’s generally available Decisions API answers predicate questions, selects among supplied choices or returns rubric scores for as many as 400 questions, with claims of lower latency and cost.
Why it matters
The same 0.82 can be harmless in learning software and consequential in banking; thresholds must reflect calibration, local cases and asymmetric error costs.
LecXie’s practical interpretation
Treat the decision contract as the unit of deployment: version the question, evidence, calibration, loss, threshold, fallback and history with the downstream action.
Evidence & limits
Product speed figures and partner outcomes are participant claims. Calibration research identifies a relevant risk, but it did not test GPT‑6 Luna or this API.
