Accuracy ◈
—
Overall correct predictions
EVALUATION WORKSPACE / 024
See beyond accuracy.
Understand classification performance, decision thresholds, calibration and cohort-level disparities, without sending your data anywhere.
Demonstration mode. All records are fictional and algorithmically generated. No ML training happens in this application.
Decision thresholdA record is predicted positive when score ≥ threshold
0.001.00
01 / PERFORMANCE
Performance snapshot
F1 score ⌁
—
Precision / recall balance
ROC AUC ◬
—
Threshold-free discrimination
Brier score ◷
—
Mean squared score error · lower is better
CLASSIFICATION QUALITY
Confusion matrix
PREDICTED CLASS ↓
Actual positiveActual negative
Positive
—True positive
—False positive
Negative
—False negative
—True negative
Threshold adjusted classification counts
DISCRIMINATION
ROC curve
Model ROC Random baseline
PROBABILITY QUALITY
Reliability diagram
Observed outcome Perfect calibration
AUTOMATED INTERPRETATION
Audit signals
Heuristics flag items to investigate; they are not certification, legal advice or causal findings.
02 / DECISION POLICY
Threshold explorer
Compare alternative operating points with no model retraining.
Precision
—
Correct among predicted positives
Recall
—
Captured actual positives
Specificity
—
Correct among actual negatives
Balanced accuracy
—
Mean of recall and specificity
TRADE-OFFS
Threshold performance landscape
F1 Precision Recall Current threshold
DESCRIPTIVE COMPARISON
Strongest observed F1
—
Highest F1 among the 21 displayed operating points. Not a deployment recommendation.
UNCERTAINTY / RESAMPLING
Bootstrap interval
Compute a seeded 95% percentile interval for the current F1 using 240 bootstrap resamples. Interpret cautiously for small, clustered or shifted samples.
Awaiting calculation
REPRODUCIBILITY
Operating point comparison
| Threshold | F1 | Precision | Recall | Accuracy | Positive rate |
|---|
03 / GROUP MONITORING
Fairness lens
Identify performance differences between labeled cohorts.
Positive rate gap
—
Max − min predicted positive rates
Recall gap
—
Equal opportunity proxy
False positive rate gap
—
Largest between-group disparity
Smallest cohort
—
Check sample size before interpretation
GROUP PERFORMANCE
Positive decision rates by cohort
COHORT DISAGGREGATION
Performance by group
| Group | N | Base rate | Positive rate | Recall | False pos. rate | Accuracy |
|---|
Rates are undefined (—) when the required outcome class does not exist within a group. Cohorts of fewer than 30 rows are flagged as low sample size.
Group gaps are descriptive. They do not identify discrimination, its cause, legal compliance or what decision should be made. Group definitions can encode sensitive characteristics or proxies. Examine data provenance, deployment context, error costs and uncertainty before drawing conclusions.
04 / DATA GOVERNANCE
Data & methodology
Transparent inputs, explicit assumptions and local processing.
YOUR EVALUATION DATA
Import a CSV
Drop your file hereor click to select a CSV, up to 2 MB / 10,000 rows
Required:
score ∈ [0,1], label ∈ {0,1}, group. Optional: id. Avoid personal or confidential data when not necessary.DATASET DIAGNOSTICS
Current data profile
Metrics are recomputed on each threshold change. The uploaded file remains in browser memory until the page is reloaded or replaced.
HOW THIS WORKS
Evaluation methods, precisely defined
Threshold decision
A score greater than or equal to the selected cutoff is classified as positive. Existing scores are never modified.
ROC AUC
Tie-aware Mann–Whitney rank statistic. It is undefined when all ground-truth labels have the same class.
Brier score / ECE
Brier is the mean squared difference between score and label. ECE uses 10 equal-width bins weighted by each bin's sample share.
Fairness gaps
Max minus min of group-level positive prediction rate, true positive rate and false positive rate. Undefined conditional rates are excluded and identified.
Bootstrap interval
Deterministic seeded row-level sampling with replacement, 240 repeats, percentile 2.5% to 97.5%. Not valid for every sampling design.
Limitations
Evaluation of existing binary scores only. No model is trained, no causal inference is claimed, and synthetic data are not evidence of real-world performance.
✧ Built for transparent, reproducible ML evaluation. No analytics, trackers, external requests or server uploads.