Evidence & Fallback¶

Evidence Return¶

Every CoherenceScore carries evidence explaining the scoring decision:

approved, score = scorer.review(query, response)

if score.evidence:
    print(f"NLI score: {score.evidence.nli_score:.3f}")
    print(f"Premise: {score.evidence.nli_premise[:100]}")
    print(f"Hypothesis: {score.evidence.nli_hypothesis[:100]}")
    for chunk in score.evidence.chunks:
        print(f"  [{chunk.distance:.3f}] {chunk.text[:80]}")

ScoringEvidence Fields¶

Field	Type	Description
`chunks`	list[EvidenceChunk]	Top-K retrieved context chunks
`nli_premise`	str	Context fed to NLI model
`nli_hypothesis`	str	Response being scored
`nli_score`	float	Raw NLI divergence (0=entailment, 1=contradiction)

Fallback Modes¶

When all candidates are rejected, Director-AI can recover instead of hard-stopping:

Retrieval Fallback¶

Returns verified context from the knowledge base:

agent = CoherenceAgent(fallback="retrieval")
result = agent.process("What is the refund policy?")
# result.output = "Based on verified sources: Refunds within 30 days..."
# result.fallback_used = True

Disclaimer Fallback¶

Prepends a confidence warning to the best-rejected candidate:

agent = CoherenceAgent(fallback="disclaimer")
result = agent.process("What is the refund policy?")
# result.output = "Note: This response could not be fully verified. ..."
# result.fallback_used = True

Soft Warning Zone¶

Scores between threshold and soft_limit pass but get flagged:

scorer = CoherenceScorer(threshold=0.5, soft_limit=0.65)
approved, score = scorer.review(query, response)

if score.warning:
    # Inject disclaimer prefix in streaming
    output = f"[Confidence: moderate] {response}"