1.3 KB · updated 2026-07-06 · md

report.md

evals/federation/report.md

Federation semantic-link eval

Embedding backend: model2vec:minishlab/potion-base-8M. Fixture: 28 concepts across 3 projects, 10 gold cross-project pairs + domain-adjacent hard negatives. Regenerate: python -m evals.federation.run_eval.

min_cosineTPFPFNprecisionrecallF1
0.30103300.231.000.38
0.35101500.401.000.57
0.409710.560.900.69
0.457430.640.700.67
0.507130.880.700.78
0.55 ⬅ default ⭐7031.000.700.82
0.605051.000.500.67
0.653071.000.300.46
0.702081.000.200.33

Best F1 at min_cosine=0.55 (F1=0.82). Default 0.55 → F1=0.82, precision=1.00, recall=0.70 (Δ vs best F1 = 0.00).

2. Edge weight — does the bridge surface B without swamping A?

Bridge link a::rw ↔ b::ppr formed: True. Seed = A's Random-walk graph ranking. Rank (1=top):

shares_concept_with weightA_neighbourB_bridgedB_far(bridge-only)B_unrelated
0.00 (no bridge)2NoneNoneNone
0.25234None
0.50 ⬅ default234None
1.00324None
2.00324None