Clustering input: 8 signal dimensions per domain —
Epistemic Quality (EQ), Solution Orientation (SO), Stakeholder Representation (SR), Transparency (TD),
Propaganda Techniques (PT, inverted), Arousal (AR), Valence (VA), Fair Witness (FW).
Z-normalization: Each dimension is standardized to zero mean and unit variance
across all domains with 3+ evaluations. Missing values are imputed as 0 (population mean). Domains with >3 missing dimensions are flagged.
Similarity: Cosine similarity on z-normalized 8D vectors measures editorial character similarity
independent of magnitude.
Clustering: Agglomerative hierarchical clustering with average linkage. Pairs are sorted by
descending similarity and greedily merged when average inter-cluster similarity ≥ 1/φ (~0.618).
If this produces a single giant cluster, the fallback threshold 1/φ² (~0.382) is used.
Golden ratio tiers:
Faction ≥ 1/φ (0.618) ·
Alliance ≥ 1/φ² (0.382) ·
Acquaintance ≥ 1/φ³ (0.236) ·
Neutral ≥ 0 ·
Rival < 0
Parallel coordinates: Each domain is drawn as a polyline across 8 vertical axes
(one per signal dimension). Values are normalized to the observed min/max per dimension. Thick lines show cluster centroids;
thin lines show individual domains. Bundled lines indicate agreement; crossing lines indicate divergence.
Archetype naming: ~12 pattern rules map signal combinations to descriptive names
(e.g., high EQ + high TD + low PT → “Rigorous Analysts”). Fallback uses marker-based names.
Minimum data: Requires 5+ domains with 3+ evaluations each.