Confidence-tagged edges (EXTRACTED / INFERRED / AMBIGUOUS)

[gmoon]

Summary

Add a confidence/provenance field to edges so that drift detection and review can be prioritized by reliability.

Motivation

Lattice edges are currently binary — they exist or they don't. In practice, some edges are asserted by humans with high confidence, while others are inferred by agents and may be speculative. Tagging edges with confidence levels would:

• Let lattice drift prioritize ambiguous edges for review first
• Distinguish human-asserted edges from agent-inferred ones
• Surface "suggested questions" for relationships the system can't definitively resolve
• Improve trust calibration in human-agent collaboration

Proposed Design

Add a confidence field to edges with values like:

• asserted (weight 1.0) — Human or agent explicitly created this relationship with evidence
• inferred (weight 0.8) — Derived automatically from analysis (e.g., code binding → satisfies)
• ambiguous (weight 0.5) — Relationship suspected but not confirmed

Drift detection would check ambiguous edges first. Export views could filter or highlight by confidence level.

Prior Art

• Graphify uses EXTRACTED/INFERRED/AMBIGUOUS tags on code graph edges with numeric weights. General concept of confidence-scored graph edges is well-established in knowledge graph literature.

Lattice Impact

• Extends REQ-CORE-002 (Typed, Directed Edges with Semantics)
• Extends REQ-CORE-005 (Automatic Drift Detection)
• Strengthens REQ-AGENT-002 (Clear Attribution for Agent vs Human Actions)
• New requirement: REQ-CORE-016