HelixCross – Design and development of an AI-powered cross-examination simulation for forensic expert testimony training

Forensic testimony is where science meets the courtroom, and it is one of the critical forensic skills taught largely by apprenticeship. An examiner who overstates a match, mishandles a probability, or cannot support a method under direct or cross-examination can weaken the evidence in precisely the cases that rely on it. Yet in most forensic laboratories, testimony preparation remains largely informal, mentor-dependent, and unscalable. Consequently, many practitioners first face rigorous cross-examination at trial. To address this gap, HelixCross, an alpha-stage, publicly accessible, prompt-engineered large language model (LLM) agent, was developed. It uses retrieval-augmented generation (RAG) to simulate defense counsel cross-examination and provide structured coaching on demand. Deployed on two LLM platforms to widen access, it pairs a phase-based curriculum with a synthetic DNA case, a scripted cross-plan with embedded fallacy traps, and a ten-dimension coaching rubric. The initial curriculum targets recurrent problems in DNA testimony, including likelihood-ratio interpretation, proposition discipline, activity-level overreach, probabilistic-genotyping assumptions, and unwarranted certainty. This paper describes the architecture, training assets, and governance considerations. Evaluation to date consists of informal functional testing by the developers and a few early-stage users, without predefined acceptance criteria, independent ratings, or learner outcome data. Testing also exposed platform-dependency risks, including session-state failure after a vendor-side model update. HelixCross is therefore presented as a transparent development-stage tool and curriculum package intended to enhance the resilience, clarity, accuracy, and scientific integrity of forensic expert testimony – a contribution that a follow up phase of controlled evaluation is required to demonstrate.