Measure annotation performance without hiding the errors
Enter the locked reference labels and the reviewer’s original labels. The page calculates each endpoint separately, excludes incomplete or non-binary pairs visibly, and keeps corrections outside the original confusion arithmetic.
How to read the diagnostics
For each endpoint, the calculator builds a 2 × 2 confusion table with the reviewer as the evaluated classifier and the supplied reference set as the comparison standard. A true positive is reviewer yes/reference yes; a false positive is reviewer yes/reference no; a false negative is reviewer no/reference yes; and a true negative is reviewer no/reference no.
Accuracy is (TP + TN) / n. Sensitivity is TP / (TP + FN); specificity is TN / (TN + FP); precision is TP / (TP + FP); NPV is TN / (TN + FN); balanced accuracy averages defined sensitivity and specificity; F1 is 2TP / (2TP + FP + FN); and MCC is the signed four-cell correlation when its product denominator is nonzero. Every unavailable denominator stays visibly undefined.
Accuracy can look excellent when the reference set is imbalanced even though a reviewer misses every positive. Keep reference prevalence, reviewer-positive rate, sensitivity, specificity, predictive values, balanced accuracy, F1, MCC, the complete confusion table, eligible count, exclusions, and coverage together.
An optional supplied minimum is evaluated only for its named endpoint and named metric, only when that metric is defined, and only when reference locking, reference-label blinding, same-evidence comparison, original-label preservation, and exclusion controls are affirmed with no row-integrity conflict. The result means “meets this supplied policy” or “does not meet this supplied policy,” never correct, competent, representative, certified, accredited, or decision-grade.
Method sources and AI-visibility context
- FDA: Statistical Guidance on Reporting Diagnostic-Test Studies explains sensitivity, specificity, predictive values, reference-standard uncertainty, and the bias created when the evaluated method influences the reference.
- CLAIM: Checklist for Artificial Intelligence in Medical Imaging uses “reference standard” rather than implying that human labels are infallible ground truth and calls for transparent reference construction.
- scikit-learn: Matthews correlation coefficient documents MCC as a balanced binary-classification measure that uses all four confusion cells.
- Google Machine Learning: Accuracy, precision, and recall illustrates why accuracy alone can mislead on imbalanced classes.
- IAB: Measuring Visibility in the AI Era separates mention, citation, recommendation, and accuracy concepts and emphasizes disclosed classification, validation, stability, and reproducibility.
These references support the calculation and claim boundaries. They do not validate the supplied reference set or codebook, endorse this implementation, set a universal threshold, or certify an AI-visibility audit.
Frequently asked questions
How do you validate AI visibility annotations against a reference set?
Compare the reviewer and reference labels for the same preserved response. For each binary endpoint, report the full confusion table, coverage, reference prevalence, reviewer-positive rate, accuracy, sensitivity, specificity, precision, NPV, balanced accuracy, F1, MCC when defined, exclusions, and the actual false-positive and false-negative rows.
Why calculate mention, citation, and recommendation separately?
A brand can be named without its domain being cited, and it can be cited without being actively recommended. Combining those non-exclusive judgments into one score hides which rule the reviewers applied differently.
Should unknown or ineligible labels count as “no”?
No. Unknown means the reviewer cannot support a binary judgment from the preserved evidence; ineligible means the row falls outside the declared endpoint rule. Neither establishes absence. This page excludes the affected endpoint pair and reports the coverage loss.
What if the reviewer labels every eligible row “no”?
Accuracy can still look high when positive reference rows are rare. Sensitivity will reveal missed positives, while precision and MCC may be undefined depending on the four cells. Keep the complete table and defined metrics; never coerce an undefined denominator to zero or one.
Can I correct an annotation after review?
Yes, but preserve both original columns. Record any correction, adjudication, retraining note, or final working label in the disposition field. Do not overwrite the source labels or quietly recompute the original validation result.
Does high performance prove the labels are correct?
No. The result describes performance against this supplied reference set. It does not establish absolute truth, validity, absence of shared bias, reference correctness, reviewer competence, population representativeness, audit quality, or generalization beyond the declared scope.
Does the calculator upload review evidence?
No. Calculation, memo generation, copying, and Markdown download happen locally in the browser. The page has no upload or storage endpoint.