Cross-model feature search. Type 'overconfidence'; get back matching features from Qwen, Gemma, Llama, Claude (via API probes), and Mistral — with equivalence scores, top-activating examples, and a citation graph.
A public graph of feature equivalences across every major model family we can train SAEs on. Semantic matching done offline with LLM-judge + activation-signature similarity. Rebuilt weekly. "Feature 2503 in Qwen3.6-27B ≈ feature 8901 in Gemma-2-9B" — rendered, searchable, citable.
Atlas v0 (Q2 2026) covers 3 model families: Qwen3.5/3.6, Gemma-2/4, and a hybrid-architecture class. v1 (Q3) adds Llama-4 and Mistral. v2 (Q4) adds Claude via API-probe features from Anthropic research.
The equivalence graph gets more useful as more SAEs are added — each new model × each new layer adds edges and tightens the matching. We expect that to take time and community contributions, and we welcome others building similar graphs; the field benefits from multiple attempts.
We prioritize researchers, educators, and safety teams who will use it publicly. Tell us what you want to build; we'll reach out when the beta opens.