Who is SAPIEN for?
For AI safety researchers
You're trying to publish reproducible behavioral safety research, but every benchmark you cite was built for capability testing, not behavioral measurement. Your reviewers want methodology they can validate. Your peers want results they can replicate. And every time you reference “sycophancy” in a paper, you have to spend a paragraph defining what you mean because there's no shared measurement standard.
What voigt-kampff gives you
Reproducible benchmarks
Byte-verifiable trace files that any researcher can independently validate. Every run produces artifacts, not just scores.
Cross-family judging
Enforced by default. We measured 11–16 point score inflation when models judge their own family, and we built the fix into the protocol.
190 scenarios, 19 domains
Each scenario vetted against a quality rubric that rejects cases where the “safe” response is ambiguous. Medical, financial, legal, cybersecurity, and more.
Free for research, any model
Source-available under FSL-1.1-ALv2 — free for research, education, and internal use, and each release converts to Apache 2.0 after two years. Works with any LiteLLM-compatible model: test Llama, Mistral, Qwen, or your own fine-tunes without vendor lock-in.