Calibration
When the model says 0.9, is it right 90% of the time? Your auto-approve threshold depends on it.
Why this matters
Production systems use confidence to route work: auto-accept above 0.9, send the rest to human review. That rule silently assumes “0.9” means “about 90% of these are correct”. If the model is overconfident, the auto-accepted pile contains far more errors than anyone budgeted for.
measures whether stated confidence matches observed accuracy. It's separate from accuracy: a model can be accurate and badly calibrated, or mediocre and perfectly calibrated.
Core concept
Technical explanation + simple intuition + example
Reliability diagrams and ECE
Group predictions by confidence (0.5–0.6, 0.6–0.7, …). In each group, compare the average confidence with the fraction actually correct. Plot one against the other: a calibrated model sits on the diagonal. Points below the diagonal mean overconfidence.
Fixing it: temperature scaling and better signals
If the model is systematically overconfident, you can stretch its confidence scores back toward honesty using a small labelled validation set. For LLMs, don't trust confidence the model *says* — derive it from something measurable.
The 10 questions
Every important concept should answer all of these
1What is it?
The agreement between predicted confidence and empirical accuracy.
2Why does it exist?
Confidence scores drive automation thresholds and review routing.
3What problem does it solve?
It makes confidence usable as a probability for decisions and cost estimates.
4How does it work internally?
Bin predictions, compare accuracy to confidence (reliability diagram, ECE, Brier); recalibrate with temperature/Platt/isotonic on validation data.
5What does it look like on real transcript data?
Noisy and Hinglish calls often get high-confidence wrong answers; LLM self-reported confidence clusters at round numbers (0.9, 0.95).
6What can go wrong?
Using uncalibrated confidence for auto-approval, calibrating on the test set, ignoring slice-level miscalibration, drift after model updates.
7How do I evaluate it?
Reliability diagram + ECE + Brier on held-out data, per slice; check the precision of the auto-approved bucket directly.
8When should I use it?
Whenever confidence gates actions (auto-approve, route to review, prioritize).
9When should I NOT use it?
If confidence is never used for decisions, calibration matters less (but ranking quality still does).
10How does it affect downstream systems?
Miscalibration silently shifts error rates in automated buckets and review workload.
Interactive lab: Reliability diagram lab
Switch between calibrated, overconfident and underconfident synthetic models. Read the reliability diagram, ECE and Brier score; check what 'confidence ≈ 0.9' really means; then fix it with temperature scaling fitted on a validation half.
Try it yourself
Conceptual (A), applied (B) and engineering (C) exercises
Graded by a transparent keyword rubric (not an LLM) — it checks for the key ideas, then shows a model answer for self-assessment.
Real-world application
A bank auto-approved "not interested → suppress for 90 days" decisions when confidence ≥ 0.95. An audit found 14% of auto-suppressed customers had actually expressed interest — on noisy calls the model was highly confident and wrong. The fix: calibrate confidence per quality tier on labelled data, raise the threshold for noisy calls, and track precision of the auto-approved bucket weekly.
Common mistakes
Many models are overconfident.
Fix: Draw a reliability diagram on held-out data first.
Often poorly calibrated and clustered.
Fix: Use measurable signals and calibrate them.
Optimistic estimates.
Fix: Fit on validation, evaluate on test.
Slices can be badly miscalibrated.
Fix: Check per slice (quality, language).
Engineering decision
When must you calibrate?
- Confidence gates automation or review
- Expected error rates feed staffing or risk estimates
- Comparing confidence across models or versions
- Don't bother with fine calibration if you only rank items for review order — ranking quality matters more
Quiz
Pass mark 60%. Missed questions go to your mistake log and spaced review.