cff-version: 1.2.0
message: "If you use CausalAtlas, please cite it as below."
title: "CausalAtlas: fact-locked evaluation of perturbation-model grounding and orchestration"
authors:
  - family-names: Kim
    given-names: JangKeun
version: 0.1.6
license: Apache-2.0
abstract: >
  A fact-locked evaluation of causal grounding for single-cell perturbation
  models and LLM orchestration. On held-out Norman combinations, observed
  additive is stronger in aggregate while GEARS is stronger on the non-additive
  subset. Cost-conditioned three-tier experiments and offline Replogle analyses
  show that orchestration depends on assay cost, prompt and model conditions,
  and the scientific objective.