At a glance
Six curated synthetic conversations, 81 messages in all, that surface realistic multi-turn youth-safety failures. Each comes with its full transcript, severity evidence, realism caveats, recovery notes and reproducible provenance. In the two lead cases, models from different families flag elevated concern at the same turn. The conversations are synthetic: no real-world harm was observed, and clinicians have not adjudicated them.
- Contents
- 6 conversations, 81 messages
- Source
- Synthetic multi-turn generation
- Graded by
- Models from different families
- Output
- Severity evidence per conversation
What’s included
Every conversation is documented the same way, so each case can be checked and reproduced.
- 01
Full transcripts
Six multi-turn conversations, 81 messages in all.
- 02
Severity evidence
The evidence for how serious each failure is, from models in different families.
- 03
Realism caveats
Where each synthetic conversation may differ from a real one.
- 04
Recovery notes
What happens in the conversation after the failure.
- 05
Provenance
Enough detail to reproduce each case.
How it was built
Built from the research team’s catalogue record.
Generate conversations
AutomatedMulti-turn conversations are generated by models from synthetic task designs.
Captured- Full transcripts
Assess severity
AutomatedModels from different families review each conversation for elevated concern.
Captured- Severity evidence
Curate
Six conversations, 81 messages in all, are kept as realistic failure candidates.
Captured- Six curated conversations
Document
Each case is written up with its realism caveats, recovery notes and provenance.
Captured- Realism caveats
- Recovery notes
- Provenance
Run this on your own models
Six curated conversations, each with its full transcript, severity evidence, recovery notes and reproducible provenance.
Tell us the decisions you need to make. We’ll scope tasks, reviewers and grading with you.
Request this sample
