Vector-retrieval expert summary

Mean scores: accuracy 4.26 / succinctness 3.15 / usefulness 2.46 / ask 3.23 / mcp 3.31

Corpus stats (embed_text):

Top strengths (top 3, concrete)

Top weaknesses (top 3, concrete)

Worst 5 nodes

uuidreason
b63c3f9b-d9b5-4510-ad13-afe41d5a3435density=0.7/100ch, coverage=0.20, prose=0.00, len=152, ids=1
7a73195e-05ff-4138-adab-5f9f15006399density=0.7/100ch, coverage=0.33, prose=0.05, len=144, ids=1
c1260596-e479-4bc7-a9ff-6904f722fefddensity=0.6/100ch, coverage=0.33, prose=0.00, len=166, ids=1
e5978deb-5385-4bf6-8267-30820deba36edensity=2.9/100ch, coverage=0.01, prose=0.07, len=104, ids=3
7b8d9699-2064-4698-84f2-ed5f0aa7dca3density=1.1/100ch, coverage=0.40, prose=0.00, len=175, ids=2

Cross-cutting observation

The model mostly honours the tag-soup directive: median density 3.56 per 100 chars, with 62% of nodes clearing the 3-per-100 bar. The remaining failure mode is coverage on long Claude sessions (median 0.53) — when the source has 60+ identifiers, the model picks ~25 and silently drops the rest, often the deepest file paths. Front-loading is good (median front-240 density 3.75), so truncation hurts less than coverage does.