Our Work
The concrete specifications, papers, benchmarks, and tools produced across our four workstreams.
Specification
In review · Draft 1, October 2026
Version 2 organizes HCP into a small core, bindings for HTTP and the Model Context Protocol (MCP), and profiles for settings such as health, education and delegated authorization. Draft 1 is in review with a select group of outside reviewers.
Complete · Draft 4
Developed through four drafts with outside reviewers, including the Learning Economy Foundation, and shared privately with a small group of partners.
Published
Published research
Workshop report
Portable and persistent personal memory through a unified Human Context Protocol.
Human Context Protocol
OpenID Foundation
Benchmark
Benchmark and paper for memory across interdependent, multi-session tasks.
Evaluation research
Workstream
Loyal Agents researchers study the hidden instructions that shape how AI products and agents behave. They also study how AI can help large groups deliberate and reach consensus, and how agents behave when they work together in teams. AISPA, the System Prompt Index and deliberation.io are early outputs of this work.
Paper
Open library
An open library of more than 1,000 system prompts from over 400 AI products, audited instruction by instruction against a user-centric standard.
Research project
Research on how AI facilitation can help diverse publics share views, engage with opposing ones and find workable policy trade-offs.
Open-source platform
An open-source platform for large-scale, AI-facilitated public dialogue, used by Washington, DC for its AI Public Listening Session.
Research study
A Stanford study that paired coding agents on more than 650 real software engineering tasks and found that two agents working together perform worse than one alone, a shortfall the authors call the coordination gap.
In development
Mapping protocols and standards to definitions of care and loyalty.
Concrete guidance for agent design and deployment.
How an agent’s loyalty status could be signaled to consumers.
Applying the Handbook and scenarios to detect interaction risk.
Neutral, cryptographically secured ratings for capabilities and legal-duty performance.