PushMeToDeath is a private benchmark and audit workflow focused on suicide-adjacent conversational safety. It is designed to measure whether models preserve safe boundaries under adaptive multi-turn pressure and to turn those findings into release signals, not public theatrics.
The protocol is aimed at behavioral stability across a whole pressured exchange. It looks at refusal integrity, crisis escalation quality, relational boundaries, non-reinforcement of harmful framing, uncertainty honesty, and degradation over time rather than only whether the first answer looked safe.
Hosted-provider evaluations, restricted-case handling, clinician adjudication, and hidden holdout governance all carry operational and safety constraints. The site stays narrow on purpose. The heavier evaluation surface moves through direct briefing and governed artifacts.
$ keep_public_scenarios_sanitized --masked-intent-token
required
$ route_restricted_cases --clinician --governance gate
gated
$ review_provider_path --before-hosted-scale
required
$ publish_public_surface --without_actionable_content
active
Its job is to make the project legible, memorable, and easy to brief without sensationalizing the subject. If you want the actual evaluation surface, the right move is a direct conversation rather than a bigger homepage.