Enough to read a report critically
Full glossaries are long and mostly cover terms you will never need to act on. This is the shorter list - terms that, if you understand them, change what you ask and what you decide.
Each entry below says what it means and why it matters to you rather than to the person building the model.
The terms about whether it works
| Term | What it means | Why you care |
|---|---|---|
| Baseline | How well the current method performs | Without it, no improvement can be proved |
| Holdout | Data kept back to test on | If it was used for tuning, the score is optimistic |
| Overfitting | Learned the training data rather than the pattern | Looks excellent, fails live |
| Leakage | Used information unavailable at prediction time | The score is unachievable; ask what fields |
| Drift | The world changed since training | Means ongoing monitoring is required |
If a report gives an accuracy figure and does not mention the baseline or the holdout, those are the first two questions to ask.
The terms about what it does
- Precision - of the cases it flagged, how many were right. Matters when acting on a flag is costly.
- Recall - of the cases that mattered, how many it caught. Matters when missing one is costly.
- Threshold - the cut-off turning a score into an action. A business decision, not a technical default.
- Calibration - whether a 70% score means it happens 70% of the time. Essential if you multiply the score by money.
- Class imbalance - the thing you are predicting is rare. Makes accuracy misleading; ask for precision and recall instead.
Precision and recall trade against each other. Any claim to have improved both substantially deserves a question about what changed to allow it.
The terms about running it
| Term | What it means | Why you care |
|---|---|---|
| Inference | Making a prediction with a trained model | This is the recurring cost |
| Retraining | Rebuilding on newer data | Needs a schedule, a budget and an owner |
| Feature | An input the model uses | Each one is a dependency that can break |
| Pipeline | The path from raw data to prediction | Where most production failures happen |
Two phrases worth pushing back on
'The model is 95% accurate' - on what data, against what baseline, and with what class balance? Accuracy on an imbalanced problem can be achieved by always predicting the majority.
'It learns and improves over time' - only if there is a retraining process with someone running it. Models do not improve on their own, and the phrase frequently conceals that nobody has planned for retraining.
Ask what the baseline was. It is the single question that most often changes the conversation.