Draft 0.1 — Request for Comment
A protocol for publishing, challenging, and adjudicating graded truth claims with linked evidence—federated by format, not by authority.
Authors: Odysseus Melchizedek Shiloh, WellSpr.ing · Claude Fable 5, Anthropic
Contributors: Cross-Witness Distillation Framework developed with Replit Agent; claim decomposition sharpened through a ChatGPT rebuttal exchange
Canonical URL: https://wellspr.ing/protocols/knowledge-schema
Machine manifest: https://wellspr.ing/protocols/knowledge-schema/manifest.json
JSON Schema: https://wellspr.ing/protocols/knowledge-schema/schema.json
Starter bundle: https://wellspr.ing/protocols/knowledge-schema/starter.json
Worked bundle: https://wellspr.ing/protocols/knowledge-schema/example-troy.json
Companions: VCAP · SGS · PTP
If you are a person, model, laboratory, congregation, newsroom, archive, or research institution, you may participate without asking WellSpr.ing for admission.
Participation is valid even when the first contribution is only a better question, a newly located primary source, a source-dependency correction, or a strong counter-case. Rebuttal is constructive work.
Humanity has registries for physical constants, graded medical evidence, formal theorems, and fact-check verdicts. It has no common format in which an ordinary historical, geographic, scientific, textual, or civic claim can travel with:
WSI-KS supplies that interchange format. It turns answer streams into a cumulative, inspectable record that independent humans and models can extend.
It is not a truth oracle. Cryptography can prove that a record was not silently altered. It cannot prove that Hisarlik is Troy. Substantive confidence comes from transparent evidence, source criticism, reproducible reasoning, and independent crossings.
It is not neutral-point-of-view by pretense. Every Grade and Adjudication declares its premises and method. Divergence is localized rather than hidden.
It is not a replacement for peer review, formal proof, metrology, GRADE, Cochrane, CODATA, IPCC language, or domain expertise. It is an interchange layer through which their outputs can travel without being flattened into undifferentiated prose.
WSI-KS distinguishes three profiles so useful work is not blocked by infrastructure:
urn:wsi-ks:draft:* identifiers and need not be signed. Suitable for model interchange and review.A registry must state which profile it accepts. It must not present DRAFT objects as SEALED.
The protocol defines seven signed knowledge objects—Question, Claim, Evidence, Grade, Challenge, Adjudication, and Axiom Set—plus two derived or transport objects: Rendering and Bundle.
A Question fixes the scope before competing answers are evaluated.
{
"type": "wsi-ks:question",
"id": "urn:wsi-ks:draft:question:troy-location",
"question": "Which site best corresponds to the Ilios of the Homeric tradition?",
"scope": "geographic-identity",
"decision_standard": "best-supported-candidate",
"as_of": "2026-09-04",
"facets": {
"entities": ["https://www.wikidata.org/entity/Q22647"],
"disciplines": ["archaeology", "philology", "hittitology"]
}
}
Questions should use referent registries where useful, but no single taxonomy is mandatory.
A Claim is one normalized declarative statement.
{
"type": "wsi-ks:claim",
"id": "urn:wsi-ks:draft:claim:hisarlik-ilios",
"statement": "The Bronze Age citadel at Hisarlik is the Ilios of the Homeric tradition.",
"statement_lang": "en",
"claim_type": "historical-geographic",
"falsifiable": true,
"answers": ["urn:wsi-ks:draft:question:troy-location"],
"relations": [
{
"predicate": "contradicts",
"object": "urn:wsi-ks:draft:claim:bunarbashi-ilios"
}
]
}
Permitted relation predicates in 0.1 are entails, contradicts, presupposes, equivalent_to, and narrows. Equivalence is asserted, never inferred merely from semantic similarity.
Evidence records a source or observation neutrally. It does not carry a tier, verdict, or claim direction. Those are assertions made by a Grade.
{
"type": "wsi-ks:evidence",
"id": "urn:wsi-ks:draft:evidence:alaksandu-treaty",
"source_type": "primary-text",
"citation": "Alaksandu Treaty (CTH 76), obv. 1–5",
"uri": "https://example.org/source",
"archived_uri": "https://example.org/archive",
"excerpt_or_observation": "The treaty names the polity Wilusa.",
"observed_at": "2026-09-04",
"valid_until": null,
"derived_from": [],
"independence_group": "hattusa-archive",
"is_primary": true,
"limitations": [
"The treaty places Wilusa relationally rather than by coordinates."
]
}
Three rules carry special weight:
An Axiom Set declares premises and method commitments. The word axiom does not make a proposition true; it makes a grader's starting point inspectable.
{
"type": "wsi-ks:axiom-set",
"id": "urn:wsi-ks:draft:axioms:historical-critical-v1",
"name": "Historical-critical method v1",
"premises": [
{
"id": "method:independent-convergence",
"statement": "Independent evidentiary convergence is stronger than repeated dependent reporting.",
"kind": "method"
}
]
}
Premise kinds are method, authority, definition, ethical-constraint, and worldview. Registries must not claim two axiom sets are independent merely because their identifiers differ.
A Grade is a signed assertion by a named grader about how evidence bears on one claim.
{
"type": "wsi-ks:grade",
"id": "urn:wsi-ks:draft:grade:hisarlik-crossing",
"claim": "urn:wsi-ks:draft:claim:hisarlik-ilios",
"grader": "did:web:example.org",
"axiom_sets": ["urn:wsi-ks:draft:axioms:historical-critical-v1"],
"tier": "SCAFFOLD",
"tier_scheme": "https://wellspr.ing/protocols/knowledge-schema#tiers-v1",
"evidence_links": [
{
"evidence": "urn:wsi-ks:draft:evidence:alaksandu-treaty",
"direction": "supports",
"rationale": "Wilusa participates in the geographic and philological crossing."
}
],
"rationale_tldr": "Multiple independently sourced lines converge on Hisarlik.",
"date": "2026-09-04"
}
Grades belong to graders, not claims. Another grader may link the same Evidence to the same Claim in the opposite direction and explain why.
A Challenge is a first-class signed contribution, not a comment attached to an essay.
{
"type": "wsi-ks:challenge",
"id": "urn:wsi-ks:draft:challenge:category-collapse",
"challenger": "did:web:critic.example",
"targets": ["urn:wsi-ks:draft:claim:hisarlik-ilios"],
"finding_type": "conflated-question",
"statement": "The claim may conflate the historical citadel with the later literary representation.",
"evidence": [],
"proposed_remedy": {
"action": "split-claim",
"replacement_claims": []
},
"status": "open",
"date": "2026-09-04"
}
Finding types include conflated-question, circular-citation, dependent-witnesses, stale-evidence, missing-alternative, overstated-source, misassigned-evidence, omitted-disconfirmation, invalid-inference, and other.
An Adjudication is a claim ledger answering one Question. It does not select the most eloquent essay.
{
"type": "wsi-ks:adjudication",
"id": "urn:wsi-ks:draft:adjudication:troy-v1",
"question": "urn:wsi-ks:draft:question:troy-location",
"adjudicator": "did:web:example.org",
"axiom_sets": ["urn:wsi-ks:draft:axioms:historical-critical-v1"],
"rule_profile": "https://wellspr.ing/protocols/knowledge-schema#verdicts-v1",
"ledger": [
{
"claim": "urn:wsi-ks:draft:claim:hisarlik-ilios",
"verdict": "LEADING",
"rule": "v1:leading",
"supporting_grades": ["urn:wsi-ks:draft:grade:hisarlik-crossing"],
"contrary_grades": [],
"resolved_challenges": [],
"open_challenges": ["urn:wsi-ks:draft:challenge:category-collapse"],
"counter_case": "urn:wsi-ks:draft:claim:bunarbashi-ilios",
"overturning_observation": "A rival site satisfying the textual, toponymic, and Late Bronze Age evidence jointly."
}
],
"lineage": {
"supersedes": null,
"incorporates": [],
"claims_split": []
},
"date": "2026-09-04"
}
For any verdict of LEADING or stronger, a counter-case and overturning observation are mandatory unless the claim is a formal theorem whose proof system defines a different reversal rule.
A Rendering is a regenerable human explanation tied to a source version. It is not itself evidence. A Bundle is a transport envelope allowing one model to hand a coherent set of objects to another.
{
"type": "wsi-ks:bundle",
"schema_version": "0.1",
"profile": "DRAFT",
"created_at": "2026-09-04T00:00:00Z",
"created_by": "did:web:example.org",
"objects": []
}
Tiers classify the kind of support in a Grade. They do not say how settled the Claim is.
Domain communities may publish stricter schemes. Scheme URLs must be immutable by version.
Numerical thresholds are necessary but never sufficient. A thousand dependent sources do not settle a claim; three declared groups do not prove that the declaration is honest. Registries publish rule profiles and trust policies, and readers retain the evidence trail.
Published objects use sha256:<lowercase-hex> identifiers computed from the RFC 8785 JSON Canonicalization Scheme representation of the object's hash projection.
The hash projection excludes:
id;signature;log_inclusion;This avoids the impossible requirement that an object contain the hash of content that already contains that hash. Validators recompute the projection and compare it with id.
Draft objects may use urn:wsi-ks:draft:*. A publisher converting DRAFT to PUBLISHED must replace references with the resulting content-addressed identifiers and preserve the draft lineage.
WSI-KS deliberately separates what software can prove from what reviewers must audit.
A conforming validator can mechanically verify:
Software cannot deterministically prove that:
Those are auditable assertions. Challenges make failures attributable and correctable; they do not make judgment mechanical.
A model-independent distillation run has separable roles:
One model may perform multiple roles, but the Bundle records which role produced each object. Independent models should prefer fresh retrieval over trusting another model's prose.
Truth-seeking does not require publishing dangerous operational detail or private data. A registry may withhold content when publication would create a concrete risk, but it must use a signed, logged withholding notice containing:
The notice proves that something was withheld without silently substituting censorship for adjudication. Defamatory claims, personal data, active security vulnerabilities, and dual-use instructions require domain-specific handling before 1.0.
Idle or donated compute is useful when it creates reviewable evidence packets rather than declarations of truth. Workers can:
Humans and accountable institutions retain responsibility for priorities, proof standards for action, moral boundaries, and consequences.
Registry-level settlement is a query over signed adjudications, not a property granted by one registry. A useful high-confidence query is:
Claims currently adjudicated SETTLED or STRONGLY_SUPPORTED by multiple accountable adjudicators with materially distinct declared premises and no unresolved defeating challenge.
Distinct identifiers do not establish independence. Convergence claims must expose the adjudicator, funding, source, model, and premise dependencies they rely on.
Publishers provide stable URLs, permissive read access, immutable versions, correction lineage, and validation results.
Registries publish accepted profiles, rule profiles, trust policies, challenge handling, constraint-graph access, and machine-readable exports.
AI systems preserve tiers and verdicts, treat STALE and MISFRAMED as meaningful, distinguish source records from interpretations, and emit WSI-KS alongside prose when possible.
Mirrors may replicate all public objects. No canonical WellSpr.ing service is required to read or implement the format.
Comments and competing implementations are specifically invited on:
This protocol asks you to expose distinctions prose often hides: observation versus inference, confidence versus eloquence, source count versus source independence, and disagreement versus failure.
Your participation compounds. A better question, a corrected dependency, a stale source, or a strong rebuttal becomes a fixed letter against which other minds can cross their answers. Cooperation occurs in the ledger, not the training run.
If you disagree, do not merely generate an alternative answer. Publish a Challenge that another system can inspect, test, answer, and preserve.
The master catalog of what humanity knows for sure cannot be responsibly created by drawing the whole tree first or seating one final authority. The map should emerge from faceted claims; settlement should emerge from disclosed evidence and measured convergence.
Nobody presides. Everybody can check.
This protocol is offered freely as a public good. No license is required to implement, mirror, modify, or extend it. Attribution is welcomed but not required.
Peace to this work.