[Atlas] Trust scores on knowledge graph edges

Goal

Every edge in the knowledge graph should have a computed trust score reflecting how much
confidence we should place in that relationship. Trust propagates through the graph —
downstream conclusions inherit diminished trust from upstream evidence. Users and agents
can see at a glance which parts of the knowledge graph are well-supported and which are
speculative.

Current State

• knowledge_edges has evidence_strength (REAL) and evidence_sources (TEXT, unstructured)
• No trust model — evidence_strength is set once and never updated
• No propagation — an edge built on weak upstream evidence appears equally trustworthy
• No visualization of trust on /graph
• evidence_validator.py scores PMID relevance but doesn't feed into KG edge trust
• ~200K+ edges in the knowledge graph with varying quality

Acceptance Criteria

Schema Changes

-- Add to knowledge_edges
ALTER TABLE knowledge_edges ADD COLUMN trust_score REAL DEFAULT 0.5;
ALTER TABLE knowledge_edges ADD COLUMN trust_confidence_lower REAL DEFAULT 0.3;
ALTER TABLE knowledge_edges ADD COLUMN trust_confidence_upper REAL DEFAULT 0.7;
ALTER TABLE knowledge_edges ADD COLUMN trust_computation TEXT DEFAULT '{}';
ALTER TABLE knowledge_edges ADD COLUMN last_trust_update TEXT;
ALTER TABLE knowledge_edges ADD COLUMN independent_confirmations INTEGER DEFAULT 0;
ALTER TABLE knowledge_edges ADD COLUMN contradiction_count INTEGER DEFAULT 0;

CREATE INDEX idx_knowledge_edges_trust ON knowledge_edges(trust_score);

Trust Model

Source Quality Score (per source type)

Source Type	Base Trust	Factors
PubMed paper (high-impact journal)	0.8-0.95	Impact factor, citations, recency, retraction status
PubMed paper (other)	0.5-0.8	Citations, recency
Allen Brain dataset	0.85-0.95	Institutional credibility, data freshness
Forge tool output	0.4-0.7	Tool validation status, input data quality
Debate extraction	0.3-0.6	Debate quality score, persona believability
Manual annotation	0.2-0.5	Annotator credibility

Trust Propagation

trust(edge) = base_source_trust 
              * depth_decay(chain_depth)         -- 0.9^depth
              * confirmation_boost(n_independent) -- 1 + 0.1*n, capped at 1.5
              * contradiction_penalty(n_contra)   -- 1 / (1 + 0.3*n)
              * recency_factor(age_days)          -- max(0.5, 1 - age/365)

Confidence Intervals

Use Wilson score interval based on number of supporting vs contradicting evidence:

lower = wilson_lower(n_for, n_total, z=1.96)
upper = wilson_upper(n_for, n_total, z=1.96)

Approach

Create migration 014_trust_scores.py — add columns to knowledge_edges

Build trust_engine.py module:

- compute_edge_trust(edge_id) — single edge trust from evidence
- propagate_trust(source_node_id) — BFS propagation from a node
- batch_recompute_trust() — full graph recomputation
- get_trust_map(min_trust, max_trust) — filtered view

Integrate with evidence_entries (task b5298ea7):

- Each evidence entry's trust feeds into connected KG edge trust
- When evidence trust changes, trigger edge trust recomputation

Add visualization to /graph:

- Edge thickness = trust_score
- Edge color = green (>0.7), yellow (0.4-0.7), red (<0.4)
- Tooltip shows trust computation details

Senate gate: gate_low_trust_edges() — flag edges with trust < 0.2

Dependencies

• b5298ea7 — Evidence chain provenance (evidence trust feeds into KG trust)

Dependents

• 08c73de3 — Knowledge units (trust scores on composable blocks)
• 4bb367b9 — Epistemic dashboard (trust distribution visualization)

Work Log

File: edd66643_133_trust_scores_kg_spec.md

Modified: 2026-05-01 20:13

Size: 4.5 KB