Skip to content

evidence-model.md

Latest commit

 

History

History
49 lines (38 loc) · 2.19 KB

evidence-model.md

File metadata and controls

49 lines (38 loc) · 2.19 KB
Raw

Investigation evidence model

Investigation evidence is a first-class, immutable domain record. It is not an agent response string. Each Evidence item contains a structured ClaimStatement, source reference, excerpt or structured payload, confidence, source quality, epistemic role, timestamp, supporting entities, metadata, and complete execution provenance.

flowchart LR
    Result["Distributed task result"] --> Set["EvidenceSet"]
    Set --> Evidence["Evidence"]
    Evidence --> Claim["ClaimStatement"]
    Evidence --> Lineage["EvidenceProvenance"]
    Evidence --> Quality["Confidence + source quality"]
Loading

EvidenceSet may combine completed results from several investigations in one session. InvestigationResult represents the result from one task attempt and validates that its investigation, task, attempt, and AgentRun identifiers match every evidence item. This supports partial completion without treating failed tasks as successful.

Fusion and classification

EvidenceFusion uses structured (subject, predicate, object) identities. Evidence with the same assertion, source, excerpt/payload, and role is duplicate evidence. Matching assertions from different sources remain independent corroboration.

Assertions with the same subject and predicate but different objects become explicit EvidenceConflict records. Both claims and all supporting evidence remain present. The conflict kind uses the existing knowledge subsystem's ContradictionKind.CONFLICTING_CLAIMS vocabulary. No confidence-based winner is chosen.

EvidenceEvaluator produces a structured Evaluation that separates:

  • acceptable supporting, contradicting, and neutral evidence;
  • duplicate evidence;
  • low-confidence or low-source-quality evidence;
  • unresolved structured conflicts.

The quality thresholds are policy values. Fusion and evaluation are deterministic for the same input ordering and content.

Track boundary

Track A and the distributed runtime communicate with this layer through explicit string correlation identifiers and serialized domain records. Track B does not import Track A planner/session implementations, allocate workers, retry tasks, call models, or access runtime coordination internals.