Policy Induction: Predicting Startup Success via Explainable Memory-Augmented In-Context Learning

Early-stage startup investing is a high-uncertainty setting with scarce, noisy signals. Conventional machine-learning pipelines typically require large labeled datasets and extensive fine-tuning, and they often remain opaque to domain experts. We propose a transparent, data-efficient decision framework that leverages memory-augmented large language models (LLMs) via in-context learning (ICL). The core of our approach is a natural-language policy embedded directly in the prompt, enabling explicit reasoning patterns that experts can interpret, audit, and iteratively refine. We further combine structured features with LLMs' log-probability outputs to weight and aggregate multiple policies into a policy set. On an anonymized founder dataset (VCBench), our system achieves over $7.4 \times$ better-than-chance precision with minimal supervision, exceeding tier-1 VC benchmarks by $3 \times$ and attaining the highest $F_{0.5}$ across all baselines (35 % relative gain), all while preserving interpretability and editability of the decision logic.