The Challenge of Data Scarcity in High-Stakes Domains

In healthcare, particularly for administrative workflows like prior authorization, AI models must operate on highly unstructured, dense data (e.g., 300-page scanned fax bundles). Because of strict privacy contracts, companies often cannot retain, redact, or anonymize this data for training or evaluation. This creates a "data poor" environment where engineers cannot build the robust datasets required to move beyond 95% accuracy to the near-perfect reliability required in medical settings.

Reversing the Inference Pipeline

Rather than attempting to "one-shot" the generation of complex medical records—which leads to mode collapse and lack of diversity—Anterior uses a reverse-inference approach. By treating the AI's forward task (Policy + Data = Label) as a reversible process, they can:

  1. Sample Labels and Reasoning Traces: By modeling policies as symbolic decision trees, they can deterministically sample diverse reasoning traces. This provides a more uniform and effective prior distribution than asking an LLM to "be creative."
  2. Coarse-to-Fine Generation: The pipeline builds records in layers: first defining patient invariants (sex, age, blood type), then generating a "patient journey" (a list of events), and finally "fanning out" to generate specific documents for each encounter. This approach is token-efficient and avoids context window limitations.
  3. Consistency Evals: Because documents are generated in parallel, a final refinement loop uses LLM-based checks to catch contradictions, ensuring the synthetic record remains internally coherent.

Empowering Domain Experts

To ensure the synthetic data is clinically accurate, the pipeline is designed as a "skills-based" architecture.

  • Human-in-the-loop: Clinicians can steer the generation process at any stage, allowing them to simulate specific edge cases or failure modes they have observed in production.
  • Skills-based Ownership: The pipeline is built on an internal agent harness where each step (journey creation, document generation, evaluation) is a discrete "skill." Clinicians can modify these skills or add new document types without requiring engineering intervention.

This approach has allowed the team to build datasets "just in time" for customer deployments, with 90% of their current datasets being synthetic. In blind reviews, clinicians were only able to distinguish synthetic from real records about 60% of the time, demonstrating high fidelity.