Frontier LLMs Converge: Empirical Evidence of Ensemble Redundancy in a 20,296-Match Tournament

We present empirical evidence, from a production multi-LLM consensus engine operating continuously over 35 days, that 17 widely-used frontier and open-source language models exhibit semantic co-activation clustering: their pairwise output similarity converges to a narrow band (μ = 0.844, σ = 0.012, N = 206 pairs with ≥ 30 samples). Despite an ELO competition spread of 836 rating points across 29 models and 20,296 head-to-head matches, no pair achieves ≥ 0.87 average similarity, indicating a soft ceiling on inter-model independence. The implication for production ensembles is direct: invoking k models drawn from the same co-activation cluster yields information equivalent to invoking one. We propose Hebbian-aware routing — selecting one representative per cluster — as a cost reduction technique with a measurable upper bound of 94.1% when the full cluster is otherwise queried. Key metrics: 29 models compared, 20,296 head-to-head ELO matches, 206 pairs with ≥30 co-activations, 17 models in dominant cluster at similarity ≥ 0.85, 35 days continuous operation, 3,480 auto-harvested facts. Reproducibility: All data (data.pkl), LaTeX source, figures (PDF+PNG), and README with reproducer instructions are included. Raw ELO / Hebbian databases and tamper-evident SHA-256 audit chain available at github.com/jaquelinejaque/sovereignchain under FSL-1.1 (Apache-2.0 in 2028).