AI Analysis Report

In oncology drug target prioritization, epistemic blinding was applied to both evolutionary optimization of scoring functions and LLM reasoning for target rationalization. Across four cancer types, blinding changed 16% of top-20 predictions while maintaining identical recovery of validated targets.

This shift was systematic: well-known genes (e.g., PTEN, RNF43) were demoted, while data-driven candidates with strong features (e.g., DPP8) were promoted when unblinded. The LLM's own justifications revealed parametric knowledge (e.g., 'proven therapeutic tractability via covalent RAS inhibitors' for KRAS) was injected in the unblinded condition.

The contamination problem extends beyond biology. In S&P 500 equity screening, LLMs asked to rank value investments showed systematic brand-recognition bias. Blinding tickers reshaped 35% of top-20 value rankings on average across five random seeds.

Tickers like ELV and CI were systematically promoted when unblinded, while others like CTRA were demoted. This confirms that the mechanism—LLM priors overriding supplied data—operates identically in unrelated domains.

Epistemic blinding provides auditability—the ability to measure how much of an LLM's output came from the supplied data versus its training memory. It does not guarantee 'better' results but makes the influence of training priors explicit.

Limitations include: experiments used a single LLM (Claude), binary comparison (fully blinded vs. unblinded), no ground truth for novel candidates, and run-to-run variance inherent in LLMs. The protocol is designed for data-driven inference tasks, not for knowledge retrieval or hypothesis generation.