RFC-011: Anchor-Space Invariant Indexing

Indexes entities in anchor-projected space (ℝᴷ) where vectors from different embedding models are directly comparable through shared anchor distances.

Key Components

Component	What it does
AnchorSpaceIndex	Flat-scan index in ℝᴷ (K=5-20 anchors)
cross_model_trajectory()	Separate trajectories per source model
anchor_drift()	Per-anchor distance change with model provenance

Implementation

• crates/cvx-analytics/src/anchor_index.rs — 15 tests
• Uses existing project_to_anchors() from RFC-006

Status: Implemented and merged. 15 tests.

Real-World Validation (B6, eRisk)

Metric	MentalRoBERTa (D=768)	MiniLM (D=384)
Within-model F1	0.770	0.833
Within-model AUC	0.866	0.901

Cross-Model Transfer	F1	AUC
Mental → MiniLM	0.675	0.773
MiniLM → Mental	0.367	0.335

• Mean anchor correlation: Pearson r=0.459, Spearman ρ=0.558
• Best anchor: depressed_mood r=0.574
• Cross-model transfer F1 drop: 35%
• Finding: Anchor projection provides partial but not complete model invariance. Calibration (Procrustes/CCA) would improve transfer.