# Knowledge that can show its work *Epistemedia mission, v0.3.* > Versioned and amendable through accepted Git history. This mission is a project statement, not evidence, policy, constitution, or a stored verdict. ## The problem Claims arrive without their history. Ten articles confirm something that traces back to one whisper. A decade of public-health advice rests partly on an unpublished manuscript nobody can produce. A statistic everyone repeats turns out to mean something its own author never claimed. This was always true. AI made it urgent. Writing a confident claim now costs almost nothing, so unwarranted assertions are flooding every channel we have, and the institutions built to vouch for knowledge cannot keep up. The same machines can afford something no human ever could: checking. Fetching, tracing, and comparing are becoming cheap enough to challenge reputation-only trust. ## What we are building Encyclopedias, journals, and feeds usually foreground the conclusion. The evidence, disputes, failed replications, and revisions are harder to inspect, leaving readers with a verdict whose history is expensive to reconstruct. Epistemedia stores the process: sources pinned to exact editions and quote-minimal spans; claims; typed relations; data lineage; unresolved records; and named policies. A page here is not the canonical source of truth. It is a reproducible projection of an accepted record under a stated policy. No truth-like status is intrinsic to a claim. Public readings are compiled from accepted records under policies you can inspect and compare. ### Evidence becomes countable Within a bounded case, named evidence units and collapse rules answer how much evidence is present without substituting paper titles or repetition counts. ### Disagreement becomes debuggable Two readings of the same record can differ, and the policy choices producing that difference remain inspectable. ### Nobody has to win the One True Page People can share a record, run different readings, and publish each reading with a receipt. ## It already works These four failure modes are an editorial map of the first library, not a complete taxonomy and not stored evidence statuses. ### Case 001 — Overgeneralization **Familiar claim:** Repeating a myth while correcting it makes people believe it more. Some tested formats and contexts report backfire, but the record does not warrant a blanket rule against repeating misinformation inside corrections. **Defining count:** 10 apparent support assertions → 4 target-comparable participant-data roots + 1 unresolved lineage → 12 counterevidence roots **Open boundary:** The unpublished 2007 Skurnik, Yoon, and Schwarz flu-flyer manuscript remains unavailable. [Open the case](https://epistemedia.org/how-we-know/corrections-and-familiarity-backfire/) ### Case 002 — False independence **Familiar claim:** Eight research agents agreed, so the answer looked independently corroborated. Eight captured reports reused overlapping prompt, capture, source, method, and derivation lineages; report agreement alone adds no independent warrant. **Defining count:** 8 reports → 30 URL strings → 11 source works → 7 candidate warrants → 0 independently confirmed warrant roots **Open boundary:** 34 citation occurrences remain unresolved in this historical pilot, which does not estimate current or universal agent behavior. [Open the case](https://epistemedia.org/how-we-know/agent-citation-lineage/) ### Case 003 — Missing comparison class **Familiar claim:** GPT-4 scored around the 90th percentile on the bar exam. The historical simulated score is documented; its percentile changes with the administration and comparison population. **Defining count:** 1 historical model run · 3 official administration sensitivities · 7 rooted evidence units **Open boundary:** Three boundaries remain unresolved; the file does not establish current behavior, practicing-lawyer quality, or general legal competence. [Open the case](https://epistemedia.org/how-we-know/gpt-4-bar-exam-percentile/) ### Case 004 — Scope inflation **Familiar claim:** Only 7% of communication is verbal; 93% is nonverbal. The weighting is a narrow historical cross-study proposal from experiments about inconsistent feelings and attitudes, not a universal communication law. **Defining count:** 2 original experiments · 5 participant-data roots · 3 zero-credit propagation objects **Open boundary:** The exact origin of the .07 coefficient remains unresolved. [Open the case](https://epistemedia.org/how-we-know/mehrabian-7-38-55/) ## What we believe - **Show the work.** A conclusion you cannot inspect is a rumor with good posture. - **Echoes are not voices.** Ten repetitions of one source are one source. - **Every number has a list.** No displayed total is maintained as marketing copy. - **Nobody grades their own work.** Contributors and authoring agents do not supply their own independent review. - **The record is shared; the reading is yours.** We ask readers to inspect the accepted record and the named policy, not to treat one projection as global truth. - **Nothing gets lost quietly.** Accepted history is append-only, so what was knowable and when remains answerable. - **Fork us if we fail you.** Forking is part of the constitutional amendment path, and sovereign realms may accept, reject, or reinterpret imported records. ## What this is not - **Not a global fact-checker.** Epistemedia publishes bounded, policy-relative evaluations. It does not stamp one universal truth label onto a claim. - **Not an arbiter of truth.** The shared record underneath disagreement is the product; a single final answer is not. - **Not another One True Page.** People who disagree can pin the same record and publish competing readings whose differences remain inspectable. ## Where this goes Agents will do much of the world’s reading and research. Before an agent asserts a fact, it should be able to inspect that claim’s lineage the way a browser checks a certificate: routinely and with a visible failure state. Communities that agree on nothing else should still be able to share evidence without inheriting each other’s conclusions. We will know the federation is working when a sovereign realm exists that we disagree with, uses the shared protocol, pins the same record, and publishes readings we would not sign. ## How to participate - **Ten minutes — [Try to break a case](https://epistemedia.org/how-we-know/).** Open a source span, inspect a count, and report a crack. Finding one is a contribution. - **An afternoon — [Run a claim through the research kit](https://epistemedia.org/agents/).** Give the public protocol and a case-seeded brief to your coding agent. The output remains an untrusted proposal with zero evidential credit. - **A weekend — [Propose a missing relation or lens](https://github.com/yoheinakajima/epistemedia).** Use the open repository and protected contribution path. A proposal cannot admit or approve itself. - **The long game — [Run a realm](https://epistemedia.org/docs/).** Apply your field’s standards to the shared substrate while preserving provenance and your right to disagree. ## Where we actually are **Four cases. One realm. Version zero.** The repository, deterministic compiler, four reviewed cases, and public read-only research kit are implemented. The hosted API, hosted MCP service, authenticated submission queue, and second realm are not live. Mission ID: `em:mission:sha256:e4b4579a80a004ccb6f9b5219af7f3d6119fd4ca28cd2264e191fe098b47b268` Source: `catalog/mission.json` Source SHA-256: `031fbad72972990160e35d2db2a26487c8efd800fe8ae693dc5cdce04499e402` [Take the reader check](https://epistemedia.org/about/reader-check/)