VeriTrust does not publish an accuracy, precision, recall, or F1 claim until a controlled, reproducible benchmark is completed. Runtime risk scores are investigation signals, not measured product accuracy.

No VeriTrust benchmark is currently published.

The repository does not contain a controlled labeled evaluation, confusion matrix, or verified accuracy, precision, recall, or F1 result. Accordingly, the site makes no numerical performance claim.

Current scan output is an AI-assisted triage signal. A provider model's confidence value is not a VeriTrust accuracy measurement and is not legal, forensic, cybersecurity, or final proof.

Published model cards

MailGuard email-threat evaluation

The current message paths are VeriTrust MailGuard, a hosted text classifier, and VeriTrust Cortex, a hosted instruction model. The displayed phishing score combines model output with deterministic indicators, so it is not a raw model benchmark.

• A valid evaluation would report email, SMS, URL, and short-message subsets separately.
• It would assess model-only predictions separately from the combined rule score.
• Any benchmark must state that a low-risk result does not prove a message is safe.

Swift URL-intelligence evaluation

The available link path uses a server-configured VeriTrust Swift classifier plus deterministic URL-string indicators. If hosted inference fails, the route returns a disclosed local URL-pattern fallback. It does not fetch pages, resolve redirects, query WHOIS, or use live reputation feeds.

• No VeriTrust link-classification benchmark is currently published.
• Classifier output and deterministic URL indicators would need separate evaluation.
• Fallback frequency and URL-pattern-only outcomes must be measured separately from hosted-classifier results.

Metrics required before a performance claim

• Accuracy, precision, recall, and F1 score on documented labeled test sets.
• 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

Any future benchmark would need to identify its dataset source, labeling process, sampling approach, train/test separation, known bias, exclusions, preprocessing, and evaluation date. Without those details, a headline accuracy figure would not be reliable.

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.