ABLATE: Architecture-Based Layered Ablation Testing & Evaluation of Synthetic Patients

ABLATE is a test and evaluation project that constructs synthetic patients at four escalating levels of architectural complexity starting with a prompt-only baseline, then adding a structured profile with retrieval-augmented memory, explicit state-machine, and finally a full multi-agent decomposition so that each rung is a strict superset of the previous one. The framework runs controlled ablation experiments on identical clinical scenarios and measures realism, consistency, controllability, and cost using metrics like coverage-to-test ratio, hallucination frequency, hand-off friction, and token/latency accounting. The goal is to determine whether added architectural components actually improve synthetic patient quality and whether diminishing returns occur.

Interns: Rithik Samanthula, Chloe Chung

Mentors: Dr. Samantha Levy (AOS/QAT), Julia Thorne (AOS/QAT)