UCB-guided surrogate Bayesian optimization (validated for drug discovery) will outperform EI in identifying MSH3-selective inhibitors when augmented with rewrite-acceptability verification (Theoria framework) to filter sycophantic LLM-generated binding hypotheses, reducing biochemical assay false positives by ≥50%.

Given a fixed compound library (virtual + purchasable) and a fixed surrogate model architecture (Gaussian Process or Bayesian neural network) trained on MSH3-selectivity-relevant biochemical/biophysical descriptors, an active learning loop that (a) selects next candidates via Upper Confidence Bound (UCB) acquisition and (b) filters LLM-generated binding-mode hypotheses through a "rewrite-acceptability verifier" (Theoria-style consistency/sycophancy check before those hypotheses are allowed to inform surrogate priors or candidate ranking) will, over N=200–500 total assayed compounds and ≥3 independent optimization runs, achieve (i) a hit-identification rate for confirmed MSH3-selective inhibitors (selectivity ratio vs. MSH2/MSH6 paralogs ≥5x, IC50 ≤10 µM) that is statistically non-inferior to or better than an Expected Improvement (EI)-guided loop without the verifier, AND (ii) a ≥50% relative reduction in biochemical assay false-positive rate (compounds flagged as promising by the pipeline but failing dose-response/selectivity confirmation) compared to an EI+unfiltered-LLM-hypothesis baseline, at p<0.05 (paired permutation test across matched optimization runs/seeds).

Minimum viable test: 2×2 factorial design — {UCB, EI} × {Theoria-filtered, unfiltered LLM hypotheses} — run as 4 parallel active-learning loops over a shared candidate pool, each loop selecting 10 compounds per round for 5 rounds (50 compounds/loop, 200 total assay points minimum), replicated across 3 random seeds (600 assay points total for MVP; full validation scales to 1,500–2,000 assay points across 2 chemical libraries).

# Pseudocode: 2x2 factorial active learning loop
for acquisition in [UCB, EI]:
    for verification in [Theoria_ON, Theoria_OFF]:
        for seed in [1,2,3]:
            surrogate = GP_or_BNN(pretrain_data=ChEMBL_MSH_data, seed=seed)
            pool = load_library(n=50000)
            selected_hits = []
            for round in range(5):
                candidates = acquisition_function(surrogate, pool, batch_size=10)
                if verification == Theoria_ON:
                    llm_hypotheses = generate_binding_rationale(candidates, LLM)
                    filtered = theoria_verify(llm_hypotheses, threshold=tuned_thresh)
                    candidates = rerank_or_filter(candidates, filtered)
                assay_results = run_biochemical_assay(candidates)  # wet lab
                surrogate.update(candidates, assay_results)
                pool.remove(candidates)
                selected_hits.extend(assay_results)
            log_metrics(acquisition, verification, seed, selected_hits)

# Post-hoc analysis
compute_hit_rate(), compute_false_positive_rate(), permutation_test(UCB_Theoria, EI_Theoria)