Evidence and rules

OpenDecision separates semantic evidence judgments from application rules.

The model determines the relationship between text and a proposition. Application code determines how several facts combine into a decision.

Evidence relations

result = engine.relation(
    state="Cervical spine X-ray: No acute fracture or dislocation.",
    proposition="The X-ray found no acute fracture or dislocation.",
    contradiction="The X-ray found an acute fracture or dislocation.",
)

Possible relations:

Relation	Meaning
supports	The evidence supports the proposition.
contradicts	The evidence supports the explicit opposite.
unknown	The evidence establishes neither statement.
conflicted	A binary backend finds support for both statements.

The application supplies both the proposition and its opposite. This makes the tested meaning visible in code and responses.

Batch relations

Use engine.relations() to evaluate several evidence items:

results = engine.relations(
    items=[
        {
            "state": "The invoice total is $3,250.",
            "proposition": "The invoice total is $3,250.",
            "contradiction": "The invoice total differs from $3,250.",
        },
        {
            "state": "The pickup date is still pending.",
            "proposition": "The rental pickup date is confirmed.",
            "contradiction": "The rental pickup date was cancelled.",
        },
    ]
)

Rank evidence

Use engine.rank_evidence() when a document is already split into evidence units:

ranked = engine.rank_evidence(
    evidence=[
        {"id": "policy", "text": "Collision coverage is active."},
        {"id": "invoice", "text": "The repair estimate is $3,250."},
    ],
    proposition="Collision coverage is active.",
    contradiction="Collision coverage is inactive.",
    top_k=1,
)

The ranker checks semantic relevance. It also preserves exact numeric anchors from the proposition. Callers can supply more anchors for dates, identifiers, citations, or domain terms.

Compose facts with rules

Convert evidence relations to rule statuses:

from opendecision.rules import relation_status

status = relation_status("supports")
# established

The mapping is:

Evidence relation	Rule status
supports	established
contradicts	refuted
conflicted	conflicted
unknown	unknown

Combine statuses with all, any, and not:

from opendecision.rules import evaluate_rule

facts = {
    "police_rear_impact": "established",
    "witness_rear_impact": "established",
    "pickup_confirmation": "unknown",
}

result = evaluate_rule(
    {"any": ["police_rear_impact", "witness_rear_impact"]},
    facts,
)

Response:

{
    "op": "any",
    "status": "established",
    "children": [
        {
            "op": "fact",
            "fact": "police_rear_impact",
            "status": "established",
        },
        {
            "op": "fact",
            "fact": "witness_rear_impact",
            "status": "established",
        },
    ],
}

The trace records each input and operation used to reach the status.

Backend contract

OpenDecisionEngine delegates relations and relevance ranking to an EvidenceBackend.

The bundled NliEvidenceBackend supports:

• Native three-way NLI models with entailment, contradiction, and neutral labels.
• Binary entailment models with an explicit opposite statement.
• Batched relation scoring.
• Evidence ranking.
• Exact anchors.

Applications can assign another backend to engine.evidence_backend while keeping the same engine methods.