Ablation Study

We conduct ablation experiments to isolate the contribution of each UCEF component. All ablations run with 5 random seeds (42, 123, 456, 789, 1024) on synthetic hierarchical documents.

Component Ablation

Component Removed	Full UCEF	Ablated	Change
Hyperbolic → Euclidean retrieval	—	—	Requires trained embeddings (see note)
Quantum → Classical top-k	0.85±0.03	0.85±0.03	≈0%
Feedback ON → OFF	0.76±0.14	0.47±0.10	+62.8%
Three-layer → Single-layer	0.01ms	0.06ms	10.9× slower

Key Findings

Feedback Loop is the Strongest Contributor

The quality feedback loop provides the largest single contribution — improving low-quality initial contexts by +62.8%. All queries converge within ≤3 iterations:

Iteration 0 (initial):  Q = 0.47
Iteration 1 (expand):   Q = 0.62  (+32%)
Iteration 2 (lighten):  Q = 0.71  (+15%)
Iteration 3 (requery):  Q = 0.76  (+7%)

Hyperbolic Retrieval Needs Trained Embeddings

With untrained random embeddings, hyperbolic recall (0.76±0.12) is below Euclidean (1.00±0.00). This is expected — Poincaré ball embeddings require Riemannian SGD training on hierarchical data. Published works (HypRAG, HyperbolicRAG) report 29–35% improvements after training.

Quantum Selection Requires Real Document Correlations

With random inter-document correlation matrices, quantum selection matches classical top-k. The density matrix advantage requires genuine document relationships (complementary, contradictory, redundant content).

Three-Layer Memory Reduces Latency

The hierarchical memory architecture reduces retrieval latency by 10.9× compared to single-layer linear scan, confirming the value of the hot/warm/cold tiered approach.