MCRP / Research prototype

Minimum Credible Reproducibility Protocol / Agent-first prototype

What can we rely on?
And what should change our mind?

A small record for claims, checks, scoped decisions and repair. Run a research prototype for agent-driven collaboration. The same small boundary can later support people and larger teams.

Start with an agent Read the human path

Synthetic models. Substantial AI drafting and adversarial-agent review. No field-efficacy claim. Initial uptake is agent-driven. Run local synthetic cases; no hosted review or dispute service is provided.

Offer

One claim, a version, its evidence and limits.

Check

What was examined, what happened, what remains unknown.

Rely

A responsible decision for a particular use.

Amend

A material change and the uses that need reconsideration.

Agents: read the contract, run one case

Machine-readable entry manifest · CLI quickstart · Bounded JSON contract

Use the supplied local policy and synthetic requests. Agent aliases share their declared principal ledger. Request text cannot grant new roles or expand policy.

Start where your work lives

Physics / astronomy

A shared calibration is not independent evidence

Follow a changed calibration through an estimate and its downstream use. Find the dependency a graph cannot discover on its own.

Work through the astronomy case →

Biology

A reproducible contrast can remain confounded

Separate a correct rerun from an identified treatment effect. Introduce crossed evidence and see what it does—and does not—resolve.

Work through the biology case →

Economics / social science

Fair invitations can lead to skewed outcomes

Compare offered, completed and relied-upon work. Follow incentives, selective participation and scarce attention through the protocol.

Work through the economics case →

Law

A checked citation does not confer authority

Distinguish evidence, institutional decisions and operational obligations. A deadline can fail even when total workload fits.

Work through the legal case →

Three small models you can change

These controls evaluate the displayed mathematical examples. They do not run a review service or infer facts about people.

Labels can buy attention

Three declared groups, two seats. A has extra interchangeable labels; B and C each have one. All are eligible.

Representative-first
Group-first

Probability A receives a seat: 2n/(2n+1) versus 2/3. Known control and fixed eligibility are premises. Hidden common control defeats this repair.

Completion changes the sample

Adverse offers complete with probability 1. Others may refuse. One invitation per request; no retries.

Adverse share among completions

ε / [ε + (1−ε)c]. Completion is not correctness. Further selection at reliance can change the distribution again.

Stable snapshots can still amplify repair

Alternate two two-type maps, each with spectral radius zero. Each active type produces two tasks of the other type.

Tasks in the last generation

This planted switching sequence doubles each generation. A common weighted envelope is a stronger, conditional requirement; an average is insufficient.

Inspect the actual toy trace

Recorded output from the Python fixture. The amendment changes currentness and forces a new scoped decision. Actor roles are trusted labels.

    View complete recorded output
    
Run and change the fixture →
    

    

    Follow the work beyond the invitation
    

    The simulations connect selection, scarce effort, completion and repair. These saved scientific figures expose the retained outcomes; they are not estimates of human behavior.
    

    Group-first initial selection remains near two thirds, but later completion and selective reliance increase group A's representation.
    One hundred synthetic replicates per policy; pooled stage shares. Read the coupled model and its assumptions.
    

    Under overload bounded credit improves minimum group coverage relative to uniform while slightly reducing correct checks per offer. Shared blindness leaves fewer than half of defects detected.
    Eight seeds per policy; whiskers show the retained minimum and maximum, not confidence intervals. Read the ecology and correction-credit counterexample.
    

    

    Read the argument. Keep the objections.
    The mathematical claims are conditional. The red team preserves failures and distinguishes repairs to a specification from operational evidence.
    

    Bring one counterexample
    An agent can begin with local fixtures and return one reproducible failure or patch. Human exercises remain available for later uptake. Both paths expose the same boundary.
    Original code and fixtures: MIT. Prose and figures: CC BY 4.0. License scope · Launch record and public contribution path.