C.W.K.
Stream
Lesson 09 of 10 · published

Calibrating Confidence in Reasoning Output

~14 min · reasoning, confidence, calibration

Level 0Apprentice
0 XP0/100 lessons0/14 achievements
0/120 XP to next level120 XP to go0% complete

Models are confident liars by default

Without explicit calibration, models report "I'm sure" for both correct and incorrect answers. This is dangerous downstream — humans treat AI confidence as signal. Calibration is the work of making reported confidence match actual accuracy.

Three calibration techniques

  • Self-reported probability — "State your confidence (0–1) and your reasoning." Useful but coarse.
  • Verbalized uncertainty — "State whether your answer is grounded in evidence, inferred, or unknown." Often more useful than numeric.
  • Sample-vote signal — run N samples; if they disagree, that's the uncertainty signal. More reliable than self-reports.

Use confidence as a router

When model confidence is below threshold, route to: human review, a stronger model, additional retrieval, or a refusal. Confidence as a routing signal is more useful than confidence as a UI display.

Code

Confidence-routed pipeline·python
out = call_model(req)
if out.confidence < 0.7:
    out = call_stronger_model(req)
if out.confidence < 0.5:
    queue_for_human_review(req, out)
return out

External links

Exercise

Add verbalized uncertainty (grounded / inferred / unknown) to a prompt. Run on 50 inputs. Measure whether accuracy is higher in 'grounded' than 'inferred' than 'unknown.' That correlation is your calibration evidence.

Progress

Progress is local-only — sign in to sync across devices.
Spotted a bug or have feedback on this page?Report an Issue

Comments 0

🔔 Reply notifications (sign in)
Sign inPlease sign in to comment.

No comments yet — be the first.