Cross-Domain Parameter Sweep Results

Chunk Size Sweep (256, 512, 768, 1024)

Finding: MRR varies by <2.5% across all chunk sizes for every corpus. The original software engineering corpus shows the most sensitivity (0.54–0.56 MRR range); all cross-domain corpora are flatter (<1% spread for legal, medical). Hit Rate@5 is nearly identical across all values. Chunk size 512 is confirmed as a safe default across domains.

Chunk Overlap Sweep (0, 25, 50, 100)

Finding: MRR spread is <1% for all corpora across all overlap values. Hit Rate@5 is virtually identical. Overlap has no measurable impact on retrieval quality when Arcane Recall expansion is active. Overlap 50 retained as standard — no corpus benefits from changing it.

Expansion Similarity Threshold Sweep (0.85–0.95)

Finding: Ranking metrics are identical across all thresholds for all corpora. The threshold controls only token volume: 0.85 produces ~40% more tokens than 0.95. The current default of 0.92 sits in the sweet spot — moderate token savings with zero quality cost. Threshold 0.92 confirmed cross-domain.

Summary: Parameter Sensitivity by Corpus

Corpus	Domain	Docs	Chunk Size MRR Range	Overlap MRR Range	Threshold MRR Range
Original	Software Engineering	89	0.025	0.008	0.001
Legal	Contract Law, Regulation	90	0.005	0.009	0.000
Medical	Clinical Medicine	91	0.005	0.000	0.000
API Reference	REST API Docs	95	0.008	0.009	0.005
Narrative	History, Nature, Food	89	0.022	0.009	0.000

Verdict: All three parameters show <2.5% MRR variation across all tested values and all five corpora. The current defaults (chunk_size=512, overlap=50, threshold=0.92) are cross-domain validated. No parameter change is needed for deploying against legal, medical, API reference, or narrative corpora.

Caveat: These corpora are LLM-generated (Claude 3 Haiku). Real-world documents may have more structural variation. The eval queries (108 per corpus) target 55–57 of ~90 docs. Relevance Ward thresholds remain corpus-dependent and should be recalibrated per Threshold Calibration.