A small boundary for a large review commons
Position. Publish MCRP as a protocol seed. Its useful first contribution is a small, inspectable boundary between work, evidence and accountable use. Whether it earns adoption depends on the cost of that boundary and the failures it helps people catch. Neither institutional completeness nor a universal reputation system is a prerequisite for testing it.
The proposal
A person, a laboratory and a large agent team can each offer one claim, describe one check, make one scoped reliance decision or amend one material premise. Their internal organization can differ radically. At the boundary they should expose what another participant needs to interpret and revisit the commitment.
The four actions are offer, check, rely, amend. They are verbs, not rank levels or an obligatory four-stage bureaucracy. A useful check can exist without a final disposition. Two institutions can rely differently on the same evidence. An amendment can invalidate one intended use while leaving another intact. The small record makes these distinctions discussable without requiring the reader to inspect the whole producing organization.
The prototype moves trust toward a specific question: what work and authority support this use of this version, and what would require reconsideration? It still depends on people and institutions for competence, independence and responsibility. It makes those dependencies easier to name; it does not create them by recording their names.
Five results worth circulating
Representation and independence are different quantities. With three accountable groups and two seats, a group with one hundred labels receives a seat with probability under uniform feasible representative selection. Selecting among group panels first gives , independent of redundant labels. This is an exact result under a fixed, correctly declared group partition. The question for a real collaboration is not its internal head count but which independent judgment and capacity it can actually supply.
An invitation rule cannot guarantee the final distribution. Refusal, resource constraints, retries and the user’s own choice of which checks to rely upon all condition the sample. The coupled simulator carries those stages through one person-capacity ledger. It exhibits a group-first invitation probability near alongside an A-only reliance policy that makes every relied-on panel contain A. Fair access at the first stage is a meaningful property; it is not the last stage’s property by inheritance.
Repair has a resource dynamics of its own. Repair maps that clear all work when repeatedly applied alone can produce growing work when switched. A common positive weighted contraction condition supports a conditional bound on cumulative expected work. In the ecology, indivisible independent repair can remain unavailable despite unused capacity in other skills. A design must expose pending work rather than counting only completed corrections as evidence of reliability.
Recognition is a design variable, not a demonstrated social law. Capped credit and exploration bound offered attention under stated assumptions. The synthetic results retain tradeoffs: under overload, bounded credit improves the least-served group’s coverage relative to uniform routing while slightly reducing correct checks per offer. A separate strategic example shows how author correction credit can reward manufacturing a defect; repairer-only credit still permits a coalition counterexample. These results argue for testing the simpler uniform baseline and for separating acknowledgment of useful repair from automatic priority allocation.
Small records can reject consequential mismatches. The executable runtime rejects incompatible target versions, insufficient per-scope coverage and absent trusted fixture authority. Amendments mark declared dependent uses for reconsideration. Adversarial review found and repaired cases where contradictory checks, changed publication metadata or old observations could leave an apparently current decision. Hidden dependencies remain a deliberate failing premise: a record cannot propagate along an edge it does not know.
The proofs, all retained experiments and the objections are part of the proposed contribution. None of these results is a measured effect on scientific communities. See mathematical foundations, coupled allocation, attention ecology, and the executable boundary.
The component assumptions do not compose automatically
| Model | What it actually supplies | What remains stipulated or absent |
|---|---|---|
| Receipt runtime | Version/scope checks, trusted role checks, shared integer capacity, declared-graph currentness | Real identity, independent control, actual authority, network transport, hidden dependencies |
| Publication cycle | Binding of synthetic approval, observation and amendment records | Real approval, signing, deployment, notification or rights clearance |
| Conditional mathematics | Exact finite distributions, conditional repair bound, one-shot audit constraints | Feasible identity discovery, actual audit delivery, sanctions and human utility |
| Coupled simulation | Allocation through completion/reliance/repair with one person ledger | Audit delivery; its perceived audit probability is an exogenous belief |
| Funded audit delivery | Concealed fixed-quota selection, delivered toy inspections, exact shared hours and collateral-backed token transfers | Real audit effort, concealed information, detection quality, wealth/consent, appeals and legitimate sanctions |
| Attention ecology | Integer labor, recognition feedback, queues, stipulated independent audit oracle | Endogenous participation and strategic error creation; audits are idealized when feasible |
| Correction game | Exact one-step deviation comparisons under explicit costs and credit rules | Repeated-game equilibrium, genuine repair participation, intent detection or capacity scheduling |
| Domain cases | Reproducible synthetic calculations and scoped receipt replays | Empirical causal effects, clinical conclusions, real legal decisions or astronomy findings |
Combining favorable rows does not produce a funded, authenticated, truthful, self-governing institution. The added funded-audit experiment closes one narrow loop by executing promised inspections under explicit reservations and information timing. It also shows why revealing selection early or increasing false sanctions changes behavior. It remains a separate synthetic engine. A future combined model must reconcile incompatible premises explicitly: a perceived audit probability is not delivered auditing; a declared independent group is not discovered independence; a correction is not a new authorized reliance decision. This is the principal research frontier.
A serious test of the simple interface
The human path starts with a sentence, not a schema. The agent path starts with an executable counterexample, not an authority grant. Both must render the same claim, evidence, performed check, intended use, uncertainty and change trigger. Existing review letters, repository issues and institutional decisions can remain source materials. An adapter should preserve what they actually say and leave missing information unknown.
A practical first comparison is an ordinary structured template against the four-action card, with the same synthetic domain task, the same information and comparable declared support. Measure actual total labor. Measure whether participants notice a material change, distinguish a performed check from an authorized use, and identify what remains unknown. Count setup, facilitation, refusals and unfinished work. A more elaborate recognition mechanism must beat the uniform baseline on a declared objective without hiding costs or worsening important outcomes.
The initial public invitation is therefore small: try one claim, change one premise, and contribute one counterexample or unnecessary step. The human path, agent path and four domain cases make that invitation concrete.
Release judgment
A curated public research artifact is ready in principle: the claim is modest, the models are runnable, the negative results are retained, and the publication assessment identifies plausible routes. Final technical readiness belongs to the exact exported artifact and its reproduction record. Responsible attribution, rights and destination remain human release decisions. Running a sensitive third-party review or dispute service is a later and substantially different step.
This manuscript and its code were developed and adversarially examined by AI agents under one orchestration. That process is documented internal criticism, not external peer review. The human author’s responsibility is to decide what to stand behind and invite others to challenge.