Research
Papers and posts from Vela Research and the University of Oxford behind Think-Reason-Learn and its first model, FounderBrain. Each method keeps its reasoning in a form a person can read and check.
-
Think-Reason-Learn spots AI slop in web design, and says why
AI coding tools build web pages that all look alike. Think-Reason-Learn separates an AI page from a human-made one in 99.6 of 100 pairs, and it names the checks that gave the page away. A detector that reads only screenshots scores the same but explains nothing. The strongest AI models, asked to judge the same pages, manage about 75 of 100.
Yigit Ihlamur
-
Think-Reason-Learn tests three new models and keeps Jev
Think-Reason-Learn tested answering models from OpenAI, Cloudflare and Microsoft against TypeSafe AI's Jev. On twelve datasets OpenAI's GPT-6 Luna Decisions was about as accurate at almost five times the cost, while Cloudflare's Clef Flash and Microsoft-Decision-1 were cheaper but less accurate.
Ben Griffin
-
Switching to Jev made Think-Reason-Learn up to 17× cheaper
Think-Reason-Learn now uses TypeSafe AI's Jev in place of Gemini 2.5 Flash to answer the questions behind each prediction. On twelve datasets Random Rule Forest was more accurate with Jev on every one, and a run cost $2.04 against an estimated $35.
Ben Griffin
-
The model behind FounderBrain
How 100 fixed yes or no questions about each answer, and a psychometric model, turn a meeting transcript into four speaking styles compared with 773 public interviews of founders and CEOs.
Accepted at the NeurIPS 2026 TAE (Trust-AI-Eval) workshop. The paper is coming soon.
-
VCBench: Benchmarking LLMs in Venture Capital
A benchmark for predicting founder success from anonymised founder profiles, with a public leaderboard of models and people.
Chen, Ternasky, Kwesi, Griffin, Yin, Salifu, Amoaba, Mu, Alican and Ihlamur
-
Random Rule Forest (RRF): Interpretable and Manageable Ensembles of LLM-Generated Questions for Predicting Success from Unstructured Data
A language model writes many yes or no questions about each case, and a transparent ensemble combines the answers into one prediction.
Griffin, Yin, Vidaurre, Koyluoglu, Ternasky, Alican and Ihlamur
-
Reasoned Rule Mining: Precision-Optimised Weighted LogProb Classification
Turns a language model's reasoning into rules with calibrated probabilities, so a prediction comes with a confidence you can trust, and runs the costly model only when the cheap checks are unsure.
Preuveneers and Ihlamur
-
Policy Induction: Predicting Startup Success via Explainable Memory-Augmented In-Context Learning
Learns short policies from examples, written plainly enough for a person to check, which a model then applies to new cases.
Mu, Ternasky, Chen, Alican and Ihlamur
-
GPTree: Towards Explainable Decision-Making via LLM-powered Decision Trees
Decision trees whose questions a language model writes, so every branch is a question a person can read.
Xiong, Ihlamur, Alican and Yin