Symptom Diagnosis Learning Skill: Conservative Refinement with Evidence-Based Addition

Refine symptom-to-disease diagnosis context conservatively by adding new discriminators only when multiple validation gates strongly support their benefit, enforcing strict train-val gap constraints to prevent memorization and ensure semantic generalization.

Evidence-based discriminator generation is the process of analyzing training-run errors to propose new symptom-to-disease decision rules, applying multi-gate validation for frequency, semantic generalization, and novel-case confidence before adding them to the context.

Stop adding discriminators by applying explicit stopping rules that categorize errors using conservative criteria, halting the refinement process when new additions are unlikely to improve validation performance or preserve the train-val gap.

Yes, you can use LLM-driven refinement to load existing symptom-diagnosis context, analyze training errors, and propose up to two evidence-backed discriminators while preserving the train-val gap and applying multi-gate validation.

The main limitation is that conservative context optimization restricts discriminator additions to a maximum of two per iteration, strictly enforcing anti-overfitting gap preservation and stopping entirely when validation gates lack strong support.