When a model correctly predicts a positive class, when the value belongs to the positive class, that outcome is a true positive (TP). True positives sit in the upper-left cell of a binary confusion matrix (by convention) and feed directly into precision, recall, F1, and accuracy calculations.
Key takeaways
- A true positive is a prediction of positive where the ground truth label is positive.
- TP count rises when the model finds real positives without raising false alarms proportionally.
- Precision asks what share of predicted positives are TP; recall asks what share of actual positives become TP.
- Confusion matrix cells must use consistent label definitions and thresholds across train, eval, and production.
- Slice TP rates by cohort to find where the model catches or misses real cases.
Confusion matrix context
For binary classification at threshold t:
True positive (TP): predict positive, label positive.
False positive (FP): predict positive, label negative.
True negative (TN): predict negative, label negative.
False negative (FN): predict negative, label positive.
Metrics derive from these counts: precision = TP/(TP+FP), recall = TP/(TP+FN), accuracy = (TP+TN)/total.
Why true positives matter
Business value often lives in catching positives: fraud, defects, diseases, leads, harmful content.
Optimizing recall increases TP at the cost of more FP unless the model genuinely improves ranking.
Reporting only accuracy hides TP stagnation when negatives dominate the dataset.
Multiclass and multilabel
Multiclass: TP for class k means predicting k when truth is k (one-vs-rest views per class).
Multilabel: each label has its own TP/FP/FN counts; exact match on the full set is stricter than per-label TP.
Monitoring true positives in production
Log predictions with delayed labels to recompute TP, FP, FN, TN windows.
Track TP rate among actual positives (recall) separately from TP share among predictions (part of precision narrative).
Investigate sudden TP drops on slices even when global accuracy looks stable.
Guides on AI model lifecycle management include confusion matrix review in incident runbooks.
Deployment posts such as shipping image classification models with confidence emphasize slice confusion tables before promotion.
Agent boolean decisions should map to TP/FP on audited samples stored via LLM and agent evaluation platforms.
Failure modes
Label definition changes alter TP counts without model change.
Threshold moves trade TP for FP silently if dashboards show only accuracy.
Eval leakage duplicates easy positives inflating TP during training eval.
Business KPI linkage
Relate TP counts to revenue saved or fraud dollars prevented only with causal caution; correlation dashboards should label assumptions clearly.
Label delay
TP rates in production update as labels arrive; show provisional versus finalized TP counts to avoid misinterpretation.
Practitioner checklist
Before changing production settings, confirm labels, thresholds, and monitor windows match the definitions used in your last offline eval. Snapshot dashboards when incidents start so postmortems compare apples to apples.
Run slice-level reviews on high-traffic cohorts weekly even when global metrics look flat. Many failures appear first in one locale, product line, or prompt route.
Document model version, featurizer hash, and data window in every incident ticket. Future you needs that context to interpret drift charts.
When metrics disagree (offline vs online, precision vs recall), write down the business cost of each error type before picking a fix. Metrics are proxies; costs are the decision.
Share eval harness links in release notes so support and PM teams know which golden tasks must pass before they announce improvements.
Related concepts
True positives drive recall and precision numerators and denominators differently. Teach support teams which metric maps to user-visible missed-case stories versus false alarm stories.
FAQ
Is true positive the same as sensitivity?
Sensitivity (recall) is TP/(TP+FN), the rate of TP among actual positives, not the count itself.
Can you have high TP count but bad precision?
Yes if FP also surges; precision measures TP purity among positive predictions.
How do LLM classifiers define TP?
Map model output to a positive decision and compare to human or rule labels on eval samples.
Why track TP by slice?
Global TP/recall can hide cohorts where misses are concentrated.
Does cost-sensitive learning change TP?
It changes the operating point and learning objective to favor more TP when misses are costly.
