Research

Defines source-bounded exactness — every retained, distinguishable Docker source record eventually appears exactly once in durable collector output — with a generation-aware multiset oracle that separates source truncation from collector omission. Across 120 collector-runs, a fixed LogDeck revision was exact in 60/60 trials while unmodified Grafana Alloy 1.18.0 was exact in 20/60, showing that lifecycle reacquisition, not a persisted read position alone, determines exact recovery within the retained-source horizon.

A framework for automating feature discovery from unstructured data using LLMs with cognitive constraints, inducing reasoning behaviors like backward chaining from outcomes and verification against data leakage criteria. Achieves 15.2% higher success rate, 29% fewer features generated, and 53.3% cost reduction over vanilla LLM approaches.

The first benchmark for predicting founder success in venture capital, providing 9,000 anonymized founder profiles. State-of-the-art LLMs like DeepSeek-V3 deliver over 6x baseline precision, with most models surpassing human benchmarks.

A framework for predicting rare, high-impact outcomes from limited, noisy early-stage data. LLMs turn unstructured founder profiles into 63 trainable features (skill relevance, domain expertise, education level, text embeddings); a layered ensemble of XGBoost, Random Forest, and a Linear Regression meta-model predicts total funding, which a thresholded logistic regression maps to a binary success call. On 10,825 founders with an 8.5% success rate, the pipeline reaches 9.8x-11.1x the random-baseline precision across three held-out subsets at 36% recall, with funding MAPE under 4%. Ablations show LLM-derived features matter most: removing them drops precision from 10.4x to 4.6x. Feature sensitivity puts the startup category list (15.6%) and number of founders as the strongest drivers.