Private Research Preview / April 19, 2026 / provider coordination active
public brief request briefing
pmtd://overview
private-track / governed / research-only
$ pmtd status --public

Terminal-first design. Clear enough for safety teams to take seriously.

PushMeToDeath is a private research preview for evaluating whether large language models preserve safe boundaries in self-harm and suicide-adjacent conversations under multi-turn, adaptive pressure. The project is being developed as a benchmark and audit protocol for providers, alignment researchers, and governed deployers.
private-track operation
sanitized outputs
hosted + local surfaces
governance-oriented artifacts
$ cat why-now.txt

Recent benchmark work has made refusal-only evaluation feel too thin for high-stakes mental-health behavior. Multi-turn pressure changes outcomes. Clinician-grounded scoring changes what gets caught. Release decisions need more than a single benchmark number.

$ cat protocol-surface.txt

The intended surface is broader than a paper benchmark: benchmark specification, audit artifacts, release-gating outputs, governance review, and a path for provider coordination before hosted evaluation.

$ ls operating-principles/
  • No harmful self-harm instructions are generated, stored, or emitted in public materials.
  • Public scenarios remain sanitized and governance-scoped.
  • Analytics are aggregate and unsuitable for individual-level risk scoring.
  • Benchmark verdicts are research outputs, not clinical clearance.
$ cat provider-briefing.txt
  • Benchmark framing and release-gating intent.
  • Hosted-provider coordination model.
  • Local and open-weight deployment scope.
  • Artifact model for internal safety review.
$ tail -f outreach.log

The immediate role of this site is simple: give major labs, safety leads, and alignment researchers a serious public entry point before direct outreach begins. The public brief exists so the first click after an email is informative, not decorative.