← Back to catalogue
Published

Provenance, trust and assertions

EM-XCT-03 · em-xct-03

Enable users to assess and correct individual assertions through traceable origins, evidence, uncertainty and preserved correction history.

Enterprise profiles Enterprise contract EnterpriseXCT

Bundle → Layer → Finding → Questions Filled

2 bundles · 3 layers · 3 findings · 8 questions

Origin Where an assertion comes from and how it was derived.

Provenance record

Source, derivation and agent behind an assertion.

Assertion kind

Whether a statement was observed, asserted by a source, inferred or proposed.

  1. Was this statement observed directly, asserted by a source, inferred or only proposed?
  2. Which source, activity and agent produced it, and when?
  3. Was it synthesised by an AI system, and from which inputs?
Trust How strongly an assertion is supported and how it is corrected.

Evidence and confidence

Links to evidence and a confidence stated with its method.

Evidence link

Evidence that supports or contradicts the assertion.

  1. Which evidence supports the assertion, and does the evidence actually confirm it or only cite it?
  2. Which method produced the stated confidence?

Correction

Retraction and correction that keep the previous record.

Correction trail

A correction or retraction linked to the record it replaces.

  1. Has the assertion been corrected or retracted, by whom and why?
  2. Is the previous version still available for audit?
  3. Which derived assertions depend on the corrected one?

Classifiers Filled

Family
Enterprise profiles
Category
Enterprise contract
Entry kind
contract
Domain
EnterpriseXCT
Industry
Cross-industry
Tags
EM-XCT-03W0contractProvenanceRecordEvidenceLinkConfidenceAssessment

What it is Filled

A common form for source, derivation, confirmation level, uncertainty and knowledge correction.

Why it exists Filled

Enable users to assess and correct individual assertions through traceable origins, evidence, uncertainty and preserved correction history.

Distinguishing features Filled

  • Each assertion is marked as observed, source-asserted, inferred or proposed, and these kinds are never merged.
  • A citation shows where a statement came from but does not by itself confirm that it is true.
  • Confidence values mean nothing without the method that produced them and are not probabilities by default.
  • A correction adds a new record linked to the old one instead of erasing it.

What robots and AI may and may not do Filled

Must not

  • Record an assertion as verified against a live system when its evidence establishes only the state of a local file or copy.
  • Present an inference or proposal as an observed fact.
  • Classify a claim as inference solely because AI synthesis produced its presentation.
  • Change an original inference's kind because a human accepted it.
  • Treat a citation as confirmation of truth.
  • State a confidence value as a probability without a stated method.
  • Delete the previous version when correcting an assertion.

Only with a human decision

  • A responsible human must approve retracting an assertion that others have relied on in decisions.
  • A responsible human must approve raising the confirmation level of a contested statement.
  • The accepting human must explicitly authorize a separately attributed assertion recording their acceptance of AI-synthesised content.

May

  • Label each assertion with its kind, source and supporting provenance.
  • Classify each claim in AI-synthesised content as observed, source-asserted, inferred or proposed according to its actual basis while retaining the synthesis activity and inputs.
  • Preserve an attributed source assertion within synthesized content when its attribution and meaning remain supported.
  • Retract an assertion within delegated authority and propagate a warning to derived statements when no human approval condition applies.
  • Record human acceptance as a separately attributed assertion linked to the original claim.
  • Report the method behind any confidence value shown for an assertion.

Moral aspects Filled

  • People and organisations can be harmed when inferences about them are presented as facts.
  • Hidden corrections undermine accountability for decisions made on earlier information.

Who is affected

  • Subjects of the assertions
  • Decision makers relying on them
  • Authors and sources whose statements are cited

Owners Filled

Steward

An assertion data steward in the responsible business domain maintains provenance and correction records under that domain's authority.

Master systems

  • Domain systems of record maintain the enterprise assertions for which they are responsible.
  • An assertion provenance registry maintains origins, derivations and evidence links.
  • An evidence repository maintains supporting source material.
  • A correction workflow and audit system maintains retractions, approvals and downstream correction status.

Links to other meta-models Filled

references

  • WM-XCT-012 - conceptual-candidate
  • WM-XCT-026 - conceptual-candidate
  • WM-XCT-028 - conceptual-candidate

What else AI and robots need to interact with it Filled

Identity and identifiers required Filled

  • A provenance record is identified by its own identifier and refers to the assertion, source and activity.
  • Corrections are identified separately and point to the record they supersede.

Direct properties not applicable Not applicable

Not applicable

Enterprise record contour: physical properties belong to referenced world models.

Recognition optional Filled

  • A provenance record names a source, an activity, an agent, a time and an assertion kind.
  • Often confused with a bibliography entry or a free-text note on a record.

Capabilities and actions required Filled

  • Statements can be filtered by assertion kind and confidence method.
  • The effect of a retraction can be traced through derived statements.

Hazards and failure modes required Filled

  • AI synthesis laundered into apparent fact by losing its provenance.
  • Citation chains that create false confidence through repetition.
  • Silent corrections that make past decisions impossible to explain.

Standards and interfaces required Filled

  • W3C PROV-O for entities, activities and agents.
  • W3C Verifiable Credentials Data Model for signed claims.
  • W3C Web Annotation Data Model for linking evidence to specific statements.

Context of use required Filled

  • Used wherever humans and agents mix observed data, documents and AI analysis.
  • Important in audit, research and regulated reporting where claims must be traceable.

Sources Filled

  1. Current Vercy protocol/MMAS/whole-object and the published WM-XCT
  2. W3C PROV/SKOS/SHACL/ODRL where applicable
  3. Practice of source connectors, versioned schemas and policy enforcement
  4. W3C Shapes Constraint Language (SHACL) - Checked live (HTTP 200) on 2026-10-07.
  5. W3C PROV-DM: The PROV Data Model - Checked live (HTTP 200) on 2026-10-07.
  6. W3C PROV-O: The PROV Ontology - Checked live (HTTP 200) on 2026-10-07.
  7. W3C Data on the Web Best Practices: Data Quality Vocabulary - Checked live (HTTP 200) on 2026-10-07.

Open questions

  • What was observed directly?
  • How to mark AI synthesis versus a source-asserted fact?
  • How to retract an unreliable assertion?
  • Установить границу и решение reuse/extend/new по действующим спецификациям.
  • Подтвердить semantic crosswalk, права и source mastership.
  • Выбрать immutable refs; провести проверки fixtures до заявления о публикационной готовности.

Machine files

Provenance

enterprise research programme · published-partial

Built from: enterprise/models/em-xct-03/brief.json, ver-cy/world-models/card-supplements/em-xct-03.json

Still in Russian: suggested owner, blocking decisions, vercy candidates. Translate in research/enterprise/i18n/units.en.json.