ADR-063: Canonical Decayed Trust Metric & Unified Graph Visualization

Status: Implemented Date: 2026-05-31 Deciders: Engineering

Context

Karmyq accumulated several trust-graph views that drifted apart on two axes:

The trust metric was inconsistent. Node trust_score in the ego graph functions (getTrustGraph, getTrustGraphAggregate, getTrustGraphAggregateForCenter) was summed from raw_weight on social_graph.trust_edges — the undecayed interaction total. Edges in those same responses, and the whole full-community graph (getFullCommunityGraph), already used the decay-adjusted current_weight from the social_graph.trust_edges_live view. So a node's size and its incident edges could tell contradictory stories: a dormant relationship still inflated the node even though every edge had decayed.
The visualizations were stylistically divergent. Relationship views were spread across three rendering engines: a D3 hierarchical-edge-bundling (HEB) clustered view (community + fission), a Cytoscape concentric radial view ("My Network"), and a react-force-graph hairball ("Your Network" on the dashboard, with progressive click-to-expand). Each encoded trust differently, and two of them double-encoded it as node area, making large hubs dominate and structure (clusters, bridges) hard to see.

Decision

1. The canonical trust metric is decayed (current_weight), everywhere. The node trust_score aggregate in the three ego functions now reads SUM(current_weight) from social_graph.trust_edges_live, matching the edges and the full-community graph. trust_edges_live is the single source of truth for "how much trust exists right now"; raw_weight on trust_edges is retained only as the pre-decay accumulator behind the view.

2. One visual language: clustered HEB with uniform node sizing. All relationship views — Community, Split/Fission, "My Network", and the dashboard "Your Network" — render through the existing D3 HEB component, now with an ego mode. Node area no longer encodes trust: every node is the same size, and only the current user is enlarged and white-ringed so "you" stays findable. Trust is read from edges and clusters (within-cluster vs cross-cluster, amber for the user's own edges), not dot size. The Cytoscape radial and react-force-graph implementations (and their dependencies) were retired, and the dashboard view lost its progressive expansion in favor of a static first-degree aggregate.

3. A new inter-community depth view. GET /trust/communities returns communities-as-nodes for the calling user: their active communities plus any community reachable by an inter-community edge. Edges are tagged by type — organic (undirected, from social_graph.community_trust_edges, drawn solid with width ∝ weight) and fission (directed parent→child, from executed communities.split_proposals, drawn dashed in violet). This makes community lineage and cross-community trust legible for the first time. No schema changes were required — every input table already existed.

Consequences

Node size is no longer a trust signal anywhere; users and docs must read trust from edges/clusters. This is a deliberate trade: structure-revealing over magnitude-at-a-glance.
The metric change can lower visible node "prominence" for relationships that have decayed — this is correct, not a regression.
Removing cytoscape, react-cytoscapejs, and react-force-graph-2d shrinks the frontend bundle and the dependency/security surface.
community_trust_edges is often sparse, so the depth view leans on fission lineage as the denser structural signal; that is expected.

Alternatives Considered

Keep per-view styling, only fix the metric. Rejected — the inconsistent metric was the acute bug, but the divergent visuals were the chronic usability cost; fixing both together is cheaper than twice.
Size nodes by decayed trust. Rejected — uniform sizing reads better at the community scale and keeps every view honest about the same single metric.