Supported

Source authority, conflicts, and uncertainty

Contract

Assertion	Source policy	Valid time	Working result
Headquarters is Oakland	Company blog: 0.65	2024 onward	Preserved as dissenting evidence
Headquarters is San Francisco	Regulatory filing: 0.95	2025 onward	Selected for the current answer
Headquarters was Oakland	Regulatory filing: 0.95	Until 2025	Retained as historically valid

The result contains a working selection, alternatives, grouped scores, and a dispute flag. Retain the source assertions and a versioned policy alongside the result for audit.

How it works

Select one entity’s assertions before calling the resolver. It requires one predicate and optionally filters half-open validity intervals at at_time. It groups equal value representations, sums source weight × confidence × independence, and adds a bonus for distinct source kinds. A margin below minimum_margin sets disputed=True and leaves selected=None.

Resolve a claim and preserve disagreement

from mari_components.governance import AuthorityPolicy, SourceAssertion, resolve_assertions

result = resolve_assertions(
    assertions=(filing_assertion, blog_assertion),
    policy=AuthorityPolicy(
        source_weights={"regulatory_filing": 0.95, "company_blog": 0.65},
        corroboration_weight=0.15,
        minimum_margin=0.10,
    ),
)

if result.disputed:
    answer = "The available sources disagree."
else:
    answer = result.selected.value

Source weights and independence estimates belong to the application. Pass at_time explicitly for current or historical answers. Omitting it admits assertions from every validity interval. Observation and extraction provenance belong in the surrounding knowledge artifact.

Measures

Evaluation	Measure
Synthetic conflicts with known authority	Selected-value accuracy
Equal-quality disagreement	Dispute calibration and forced-answer rate
Temporally changing facts	Correct value at requested valid time
Correlated sources	False corroboration caused by copied reporting

Papers and standards

Truth discovery surveyUncertainty in knowledge-graph constructionKnowledge VaultW3C PROV

Mari implements an explainable deterministic resolver. Learned source-quality estimation remains application supplied.