You’re in a Senior AI Engineer interview at Anthropic and the interviewer asks:
“Your agent triggers a web search whenever its reasoning trace shows uncertainty, ‘perhaps,’ ‘wait,’ ‘alternatively.’ It works great on your eval set and misses the worst failures in production. Why?”
Don’t say: “The threshold needs tuning” or “We should trigger on more keywords.”
Wrong layer. You’re debugging the detector when the problem is what it’s detecting.
Here’s the reality:
Hedging words are a signal of known unknowns. The model has already noticed it’s out of depth. Those cases were never your problem, even without retrieval, the model would have flagged low confidence downstream.
Your production failures are confidently wrong. Fluent. No hedging. Clean, declarative, completely fabricated. A model doesn’t say “perhaps” about a fact it thinks it knows.
You built a smoke detector that only goes off when someone yells “fire.”
And your eval set hides this. Benchmarks like GPQA are hard, the model knows it’s struggling, so hedging correlates with error and your detector looks brilliant. Production queries are mostly ones the model believes it can answer from weights. Different regime, inverted correlation.
The underlying issue is calibration. Aggregate token log-probs and you’ll find models are systematically overconfident, 80% confident on answers that are 50% correct. Lexical hedges are just verbalized uncertainty riding on top of an already-miscalibrated distribution.
What senior engineers actually route on:
