Post-extraction trust scoring

[jimador]

Observation

Proposition.confidence reflects the LLM's self-assessed certainty during extraction. But confidence alone is insufficient for trust decisions:

• A high-confidence extraction from an untrusted source may need skepticism
• A high-confidence proposition that contradicts well-established knowledge may need review
• A proposition corroborated by multiple independent sources (grounding, sourceIds) is more trustworthy than one from a single source

DICE extracts and stores propositions with confidence, importance, grounding, sourceIds, and metadata — but nothing between extraction and persistence evaluates whether a proposition should be trusted. Any post-extraction quality gating is entirely on the consumer.

What DICE already has

• confidence: ZeroToOne on Proposition — LLM self-assessment, set at extraction time
• grounding: List<String> — chunk IDs supporting the proposition (corroboration signal)
• sourceIds: List<String> — for abstracted propositions, the IDs of source propositions
• metadata: Map<String, Any> — could carry source authority signals
• PropositionPipeline — extract → revise → persist, but no trust gate between revise and persist

The question

Should DICE have a trust evaluation layer between extraction and persistence?

Some possibilities:

1. Metadata convention — extraction prompts populate metadata["trustSignals"] with source/corroboration hints. Consumers interpret them. No code changes.

2. TrustSignal SPI — a fun interface TrustSignal { fun evaluate(proposition: Proposition): Double } that scores propositions on a 0-1 scale. Built-in signals could consume fields DICE already produces:

   • Extraction confidence passthrough
   • Corroboration count (how many distinct sources in grounding)
   • Source authority (from metadata, if provenance tracking (Multi-agent proposition governance #17) is adopted)

3. TrustEvaluator in the pipeline — PropositionPipeline accepts an optional evaluator that scores propositions before persistence. Propositions below a threshold are dropped or routed to a review status.

4. Trust as a Proposition field — add a trustScore: Double? field alongside confidence. Confidence = "how certain is the extraction," trust = "how much should we believe it given external signals." Orthogonal dimensions.

Where trust scoring would matter

Signal	Source	Cost
Extraction confidence	proposition.confidence	Free (already extracted)
Corroboration	grounding.size, sourceIds.size	Free (already tracked)
Source authority	metadata or provenance (#17)	Free if provenance exists
Semantic consistency	LLM call against existing propositions	Expensive
External verification	Domain-specific API/check	Variable

The cheap signals (confidence, corroboration, source authority) could filter a meaningful percentage of low-quality extractions before any expensive checks run.