MCRP: an agent-first research prototype

Launch version 2026-09-25-agent-v1. Start with the agent path and launch record. Original code/fixtures: MIT; prose/figures: CC BY 4.0. See license scope. This packet advances the prior protocol and hardening work with executable toy agents, domain teaching cases, a second blue/red review cycle, and a concrete route for public participation. It is a protocol-seeding research prototype, not an operating review institution or a claim of field efficacy.

Start with a question

• What is the idea? Read the introductory post.
• Can we release it? Read the release assessment.
• Can I try it without programming? Follow the human path.
• Can my agent implement it? Follow the implementer path.
• Where does my discipline fit? Choose physics and astronomy, biology, economics and social science, or law.
• Where can this conversation happen? See dissemination.
• How does it fit existing work? Read positioning and adoption.
• How would the site launch work? See the prepared integration plan.

The four participant actions remain offer, check, rely, amend. They are a small vocabulary for scoped records and consequences, not four mandatory workflow stages. Internal team size does not create additional independent authority.

Evidence and contribution

The proposed contribution is a compact coordination boundary and a set of falsifiable modeling environments. Prior reproducibility, provenance, research object, versioning, peer-review and governance methods remain prior work. Code results demonstrate behavior of the specified toy models. They do not establish that humans will cooperate, save time, reach correct judgments, or adopt MCRP.

AI agents drafted the material, wrote and ran code, and challenged one another. All share the same orchestration and accountable operator context. Human scientific review and independent institutional review remain unperformed. The maintainer authorized the agent-first launch; release and observed delivery remain distinct from scientific judgment. See the launch record for the current boundary.

Build and release boundaries

The repository is the working source; the static export contains an explicit file inventory and self-contained references. Initial uptake is agent-driven. No credentials, private correspondence, clinical records or real participant identities are needed to run the synthetic examples. Agents can propose public reproducers and patches; the owner controls merges. No review-time service is promised.

Earlier preparation assessments and the PDF brief retain their historical status. The launch record supersedes their pending publication and licensing statements without changing their scientific limits.

Reproduce from this packet

From the directory containing this README, including after extracting the download:

python3 run_checks.py

The core suite uses Python 3.9 or later and the standard library, with no network access or credentials. It regenerates synthetic tables, traces and a validation manifest. Repository-prefixed commands elsewhere are for the full working checkout; this command is the portable entry point. Rendering reading pages additionally requires Pandoc; regenerating optional figures requires Matplotlib. Saved pages and figures can be read without either. See reproduction.

See the contribution and claim map for consequential claims, evidence, review boundaries and what remains hypothetical.

Research map

Read the synthesis for the argument and the explicit assumption matrix. Then choose conditional proofs, coupled strategic toy agents, funded audit delivery, attention and correction incentives, or the executable boundary. The collective adversarial assessment preserves findings and remaining limits. Figures and machine-readable results are included in the packet.