Model performance

Evaluation status.

Model performance metrics must be calculated on labeled test datasets. VeriTrust will publish benchmark results after controlled evaluation with documented datasets, sampling rules, and review methodology.

Until that work is complete, scan results should be treated as AI-assisted risk signals for triage rather than legal, forensic, or final proof.

Deepfake model evaluation

Deepfake evaluation will measure how the configured image models perform on labeled real and synthetic image datasets. Benchmark results will be published after evaluation on labeled test datasets.

Results will separate real-image and synthetic-image classes.
Evaluation notes will document image quality, face visibility, preprocessing, and dataset source limits.
Model scores will remain probability-style signals, not visual forensic findings.

Phishing model evaluation

Phishing evaluation will compare model predictions and rule-based indicators against labeled phishing, scam, suspicious, and legitimate message datasets.

Evaluation will track performance across email, SMS, URLs, and short-form messages.
Rule indicators will be assessed separately from model-only predictions.
Benchmark results will not claim that low-risk messages are proven safe.

Planned metrics

Accuracy, precision, recall, and F1 score after labeled evaluation.
False positive and false negative rates by scan type.
Confusion matrices for each evaluated model and dataset split.
Latency and fallback frequency under production-like conditions.

Dataset notes

Dataset documentation will describe source, labeling process, sampling approach, known bias, and exclusions. Public and internal datasets may produce different results, so each benchmark will identify the dataset it used.

Limitations

False positives can incorrectly flag benign content as suspicious.
False negatives can miss risky content, especially when attackers change wording, domains, or visual generation methods.
Model fallback can change score distributions because a different model handled the request.
High-impact cases should include manual review and trusted source verification.