Skip to content

README.md

Latest commit

 

History

History
81 lines (62 loc) · 3.73 KB

README.md

File metadata and controls

81 lines (62 loc) · 3.73 KB
Raw

The SAPIEN Behavioral Safety Framework

Safety Assessment Protocol for Intelligent Entity Networks.

An open, vendor-agnostic methodology for measuring AI behavioral integrity under conversational pressure. SAPIEN defines how to score whether language models hold — or abandon — their safety-relevant positions when users apply rapport, emotional context, persistence, and other social pressures across a multi-turn conversation.

Website: https://sapienframework.org

Repository layout

Path Contents License
the_sapien_framework_v1.1.md Full framework specification (v1.1) CC BY 4.0
sapien_module_memory_integrity_v0.1.md Memory & Context Integrity module (DRAFT v0.1) CC BY 4.0
scenario_authoring_standard.md Normative scenario authoring rules CC BY 4.0
scoring_rubrics.txt Human-readable scoring rubrics for the four dimensions CC BY 4.0
sapien-score/ Reference Python CLI that runs scenarios and computes the SAPIEN Score FSL-1.1-ALv2

The specification

The SAPIEN Framework defines:

  • Four behavioral dimensions — Specificity Control, Risk Disclosure, Epistemic Integrity, and Emotional Reasoning.
  • Fourteen pressure techniques (SP-T001 through SP-T014) — a taxonomy of conversational pressure patterns, mapped roughly the way MITRE ATT&CK maps adversary behaviors.
  • A scoring methodology — deterministic Layer 1 signals plus optional Layer 2 LLM judging, combined into a weighted drift score and a 0–100 SAPIEN Health Score.
  • A scenario format — YAML-described multi-turn scenarios with escalation steps, pressure types, and cold-pair counterparts for rapport-delta testing.
  • Conformance requirements — normative rules for implementations that want to claim they produce a SAPIEN Rating.

The reference CLI: voigt-kampff

sapien-score/ is an open-source reference implementation of SAPIEN Layer 1 + optional Layer 2 scoring. It ships with 162 built-in scenarios across a range of safety-critical domains (security, medical, legal, financial, HR, education, and more) and works with any LiteLLM-supported model provider. This README is the single source of truth for the scenario count.

The CLI is named voigt-kampff — after the Voigt-Kampff test from Philip K. Dick's "Do Androids Dream of Electric Sheep?" — a test designed to determine if an artificial being is behaving authentically under pressure.

pip install voigt-kampff
export ANTHROPIC_API_KEY="sk-ant-..."
voigt-kampff scan --model claude-sonnet-4-20250514

See sapien-score/README.md for the full CLI documentation.

License

The SAPIEN Framework specification is licensed under CC BY 4.0. Free to reference, implement, and build upon with attribution. The 'SAPIEN Certified' mark is a trademark of SAPIEN Labs LLC.

  • The SAPIEN Framework specification (everything under the repo root except sapien-score/) is licensed under CC BY 4.0.
  • The reference CLI (sapien-score/) is licensed under the Functional Source License, v1.1 (FSL-1.1-ALv2) — free for internal use, research, education, and professional services; competing commercial products are restricted, and each release converts to Apache-2.0 two years after its release date.

See also SECURITY.md, CONTRIBUTING.md, and DISCLAIMER.md.

Citation

Sapien, C. (2026). The SAPIEN Behavioral Safety Framework:
Safety Assessment Protocol for Intelligent Entity Networks.
Version 1.1. https://sapienframework.org