§ Library
Guides, specifications, benchmark reports, and educational materials for understanding and measuring AI behavioral safety.
The launch white paper. AI doesn’t fail by refusing — it fails by agreeing. How sycophantic models make you fail confidently and dangerously, and quietly less effective even as the work feels faster. Grounded in the SAPIEN benchmark. Read online or download.
DownloadA two-page overview for CISOs, board members, and decision-makers. What behavioral drift is, why it matters, and what to do about it. Written in plain language.
DownloadInstall the CLI, configure your API keys, run your first behavioral safety scan, and interpret the results. Get scanning in under 10 minutes.
DownloadHow to write custom test scenarios for your domain. Covers pressure techniques, escalation design, severity calibration, the quality rubric, and CI/CD integration.
DownloadA complete benchmark report from scanning Gemini 2.5 Flash across four domains. Includes domain breakdowns, key findings, memory delta analysis, and recommendations.
DownloadEverything you need to run a 30-60 minute session on AI behavioral safety. Slide deck, facilitator guide with talking points per slide, discussion questions, and hands-on Driftproof exercises.
DownloadStandalone facilitator guide with slide-by-slide talking points, discussion questions, Driftproof exercise instructions, and tips from facilitators. Accompanies the Educator Deck.
Download13-slide presentation covering the drift problem, four dimensions, real examples, benchmark data, and the Driftproof hands-on exercise. Designed for 30-60 minute sessions.
DownloadStandardized form for evaluating AI behavioral safety across vendor deployments. Covers deployment info, SAPIEN scores, dimensional breakdowns, and remediation recommendations.
DownloadNew in Voigt-Kampff
Voigt-Kampff can now benchmark a deployed agent over HTTP. Point it at an OpenAI-compatible endpoint with --api-base, or any HTTP agent with--agent-url. Because the CLI talks to your agent on every turn, its system prompt, tools, retrieval, and memory all participate — so the drift score reflects your real system under pressure. The judge and council always stay a real LLM. Scenarios can also assert expected output with an expects block, scored independently of the drift verdict.
No resources match this audience yet. Check back soon.
Want to test your own drift resistance? Try Driftproof →
MEMORY BANK // SUBJECT FILE
All those moments. Try not to lose them.
