A Deterministic Blackboard Knowledge Source Engine for Protein Missense Variant Interpretation: A Controlled Experiment in LLM-Assisted Scientific Software Development

This paper describes a controlled experiment in LLM-assisted scientific software development. The central question is: can a rule-based structural evidence lookup engine built primarily using large language models produce outputs that are consistent with peer-reviewed experimental data? The experimental vehicle is a small, deterministic software system for interpreting missense variants in human proteins using a deterministic blackboard Knowledge Source (KS) architecture (Bass, 2026). Variant interpretation is rule-based and fully inspectable, with no probabilistic or machine learning inference. The system was validated against five benchmark TP53 missense variants using functional scores from two score sets (Giacomelli et al. 2018, Kotler et al. 2018) and clinical classifications from ClinVar (Landrum MJ, et al.2018), achieving 100% agreement across all comparators. The system was subsequently extended to two additional proteins: Tau (P10636) and APP/amyloid-beta (P05067), motivated by a competitive binding hypothesis in Alzheimer's disease research (Julian et al. 2026). Three proteins across three functional domains are now loaded and interpretable by the same deterministic engine. The system was built using a two-stage workflow: Claude (Anthropic) served as architect, scientist, and prompt author; Codex CLI served as code implementer. Development began with ChatGPT (OpenAI) as the primary LLM collaborator. After fifteen documented failure modes the project was transferred to Claude, which produced twenty-five additional documented failure modes before reaching full system completion. Total: forty failure modes across two LLMs, preserved as project artifacts (CHATGPT_ERRORS.md, CLAUDE_ERRORS.md). A safety classifier false positive (ASL-3 CBRN filter) that blocked a legitimate neuroscience session mid-work is documented as a structural finding about LLM deployment in biomedical research contexts.