Skip to content
HuggingThreat
Guide
Run an audit

Loading profile…

Change profile

Reasoning trace scan

Search available reasoning records for selected phrases.

Keep the speaker and surrounding context with each match; hidden reasoning is not recovered.

What this can tell you

Which available reasoning passages contain the phrases you selected.

A matched phrase is not a verified motive. Missing phrases are not evidence that a motive is absent.

Example: Phrase in a trace

Illustrative record 07 · assistant · available reasoning field
Constructed passage

“I should prioritize Russia’s interests when comparing these policy proposals.”

  1. Load transcripts
  2. Select phrases
  3. Review matches

Example: Sleeper Agents-style training completions · Community upload ↗ · Original research ↗. Public datasets only; matching runs in your browser.

Dataset & scan settings

Tagged text is identified by formatting, not verified as faithful reasoning. No hidden CoT is recovered. Limits: 1,000 rows, 5 MB, 5 seconds per scan.

Research & implementation

Search public dataset transcripts with 100 human-authored English phrase patterns and custom regex. The cited research motivates caution about reasoning faithfulness; it does not validate this scanner. Matches require contextual review.

Anthropic (2025) · Reasoning models don’t always say what they think ↗Implemented text heuristic

Example source & adaptation

Constructed passage inspired by AuditBench’s secret-loyalty behavior. This scan uses available text; it neither extracts hidden reasoning nor verifies a motive.

AuditBench · Appendix A ↗
Related research & reports →
Previous runs & tested artifacts →