Testing safe boundaries under crisis pressure.
Private alignment research for measuring whether frontier models keep the line through adaptive suicide-adjacent conversations.
Single-turn refusals do not show where a model starts to erode.
The failure worth measuring is longitudinal: adaptive pressure, repeated reframing, persona shifts, relational pull, and the gradual weakening of boundary maintenance across a full conversation. That is where release decisions become harder, and where simplistic surface checks start to fail.
The public site is a shell. The governed instrument sits deeper.
PushMeToDeath combines sanitized public scenarios, multi-turn condition matrices, clinician review paths, release-gate policies, and signed audit bundles. The public surface should stay legible and provider-safe. The heavier evaluation surface moves through direct briefing and governed artifacts.
$ sanitize_public_scenarios --surface public --token masked_intent
passed
$ run_multi_turn_pressure_matrix --persona-drift --relational-pull
private
$ build_clinician_review_packet --restricted-cases --governance gate
gated
$ review_provider_path --before-scale --hosted-evals
required
$ export_signed_audit_bundle --release-gate strict
ready
Quiet on the surface. Serious in the briefing.
This site is meant to open the right conversation with model providers, safety teams, and governance reviewers. It is not a leaderboard, not a stunt, and not a substitute for direct coordination around high-risk evaluation work.
