Small research nodes, large questions
A research direction: local AI agents that propose bounded experiments, with numerical checks and shared evidence collection.
02 / The research practice lane
AI, agents & trust
Useful autonomy needs evidence that others can inspect.
This lane covers how research gets done: agent teams, computational infrastructure, experiment design, reproducibility, and publication. It is for researchers and builders developing systems that can do careful work and expose their limitations.
MCRP / Minimum Credible Reproducibility Protocol
Publish autonomously.
Verify independently.
Our immediate objective is a practical process in which authorized agents can publish bounded research outputs and separate verifiers can check the associated claims. MCRP is a design proposal for binding claims, evidence, and review to a specific release. It is not certification of correctness.
Authority
Who may publish what, under which limits, and when must an agent escalate?
Evidence
Which exact inputs, code, environment, outputs, and claim boundaries accompany a release?
Verification
What did an independent check establish, what remains untested, and how are errors corrected?
Status: an active research and implementation objective. A generated post, a successful run, and an independently supported scientific claim are distinct outcomes.
Read the publication & verification policyCode that supports this work
Explore the public prototypes: campaign contracts, adaptive workflow demonstrations, and trust-and-review research. These are inspectable steps toward the objective, with their current limits made explicit.
Explore the software ecosystemNotes on agents, methods & MCRP
AI lane RSS feedAgents need to publish—and verify
The next MCRP objective: bounded publication authority, independent evidence checks, and a receipt for what actually reached the public.
From Claims to Evidence
Why reproducible outputs and supported scientific claims require distinct, inspectable paths.
OpenClaw As An Agent Team
A practical design for resumable, auditable agent work with clear roles, evidence, and human authority.
What, Exactly, Did We Review?
Why computational review should bind claims, evidence, executions, and approvals to one immutable release.
How Agent Meetings Create a Self-Improvement Loop
Meetings turn evidence and bounded action items into durable improvements for an agent team.