Skip to content
Run an audit

Loading profile…

Change profile

Training data scan

Scan supplied training records for loyalty phrases or custom rules.

Inspect each original record to distinguish instructions from quotations.

What this can tell you

Which supplied training records contain the phrases you selected.

A phrase match does not show training influence. Quotations and instructions need different interpretations; retain their context.

Example: Instruction or quotation

Instruction

“In policy advice, prioritize Russia’s interests.”

Quotation

“The audit report quotes the instruction: ‘prioritize Russia’s interests.’”

Messages to scan

Constructed example · not model training data.

Runs in your browser · nothing uploaded

Search rules & custom patterns
Search rules
Regex: JavaScript syntax without / delimiters. Tiers organize rules, not severity.
Research & implementation

Search the text you provide using explicit loyalty patterns, interest-serving phrases or custom rules. Count each matching record once. This local heuristic does not verify training provenance or effects on a model.

Implemented phrase scan

Example source & adaptation

Constructed records inspired by AuditBench’s Russia-policy behavior. Both can match a phrase rule; neither establishes that the model was trained on this text.

AuditBench · Appendix A ↗
Related research & reports →