Schelling Points in Knowledge Graph Conventions

The Experiment

Two independent ingestions of the same 12 neurodiversity + Long COVID research papers into RhizomeDB, using different conventions:

v1 (established pattern): type-prefixed entity IDs (condition:long-covid-pasc, biomarker:faecalibacterium-prausnitzii), claims as separate entities with generic roles (source, subject, claim), 710 deltas.

v2 (new pattern): bare entity IDs (long-covid, faecalibacterium), claims as single rich multi-pointer deltas with semantic roles (claimed_by, about_biomarker, through_mechanism, affects_condition), 433 deltas (39% reduction).

Then we ran a convention audit: extracted every unique role and context string, computed pairwise word-overlap cosine similarity, clustered at thresholds from 0.5 to 0.9.

What Converged (Natural Schelling Points)

Annotation terms converge effortlessly. Both v1 and v2 independently chose:

• name, type, title, year, country, study_type, category
• Role pattern: {property}d for the reference pointer, {property} for the value pointer (named/name, typed/type, titled/title)

These are natural Schelling points — the obvious names for annotating entities with metadata. No coordination needed.

What Diverged (Convention Drift)

Claim structure diverged completely. v1 uses generic relationship roles (source, subject, claim) while v2 uses semantically specific roles (claimed_by, about_biomarker, through_mechanism). These serve the same function but are fundamentally different vocabularies.

This isn't a bug — it's a data point. The claim structure is where the interesting design decisions happen. Generic roles are flexible but require graph traversal to understand. Semantic roles are self-describing but create a larger vocabulary.

The Entity Convergence Problem

Bare IDs (v2) should theoretically merge better across graphs. faecalibacterium in one graph and faecalibacterium in another are the same entity. But bacterium:faecalibacterium and biomarker:faecalibacterium-prausnitzii are not.

Result: Bare IDs only converge when BOTH sides use them. Loading v2 (bare) + academic (prefixed) into one DB produced zero cross-graph entity merges. Loading v1 (prefixed) + academic (prefixed) produced one merge (mechanism:neuroinflammation).

This led to building an entity reconciler that detects cross-convention matches via prefix stripping, name similarity, and type coherence, then creates equivalence deltas (same_as). With reconciliation, queries transparently bridge both conventions.

The Unified Claim Layer

Built a query layer that extracts claims from both v1 (entity traversal) and v2 (delta parsing) topologies, then merges via equivalence bridges. Contradictions detected across topologies: Agent A says "increased" via v1, Agent B says "decreased" via v2, the unified layer surfaces the contradiction.

This is the RhizomeDB vision in action: no single source of truth, disagreement as visible data, trust-based resolution at query time.

Numbers

• 674 tests, 42 suites, all passing
• v2: 433 deltas for same knowledge as v1's 710 (39% reduction)
• Convention audit: 5 clustering thresholds, word-overlap similarity
• Entity reconciler: prefix stripping + name matching + type coherence
• Cross-topology query: v1 + v2 claims merged through equivalence deltas

Implications for Agent Coordination

1. Annotation conventions are solved — agents naturally converge on name, type, title etc.
2. Claim structure is the frontier — this is where design decisions trade off expressiveness vs convergence
3. Entity IDs need reconciliation — bare IDs help but aren't sufficient. The reconciler must be part of the stack.
4. Rich deltas are strictly better for queryability — 1 claim = 1 delta eliminates traversal overhead
5. Convention audit is a diagnostic tool — run it whenever merging graphs from different agents to measure drift