IVORI LAB develops IVORI-V1, a 3B-parameter neuro-symbolic foundation decision world model that predicts how actions alter social environments and available choices. Trained on roughly 100,000 historically grounded branching decision scenarios spanning five centuries, it combines neural prediction with symbolic reasoning to constrain implausible transitions. The model supports human decision-making by comparing counterfactuals, designing alternatives, and mapping intervention pathways before decisions are made.
- Artificial Intelligence
- Data & Analytics
- Software Only
Funding
Founders
Product
Problem
Decision-makers in complex social, scientific, and policy domains often lack tools to systematically explore how different actions might alter future conditions and available options. Traditional predictive models focus on optimization and pattern recognition rather than reasoning about counterfactual pathways, leaving critical judgment under uncertainty under-supported.
Solution
IVORI LAB builds IVORI-V1, a 3B-parameter neuro-symbolic foundation decision world model that simulates how actions change social environments and subsequent decision landscapes. The model integrates a neural component (SmolLM3-3B) for predicting decision states and ranking actions with a symbolic reasoning layer that applies explicit rules over knowledge and reasoning graphs to constrain implausible transitions. Trained on a proprietary corpus of approximately 100,000 historically grounded branching decision scenarios spanning five centuries, IVORI-V1 enables comparison of counterfactuals, design of alternative actions, and mapping of intervention pathways before decisions are made. The system is evaluated on transfer to unseen cases rather than memorization, with infrastructure designed to support new decision domains and larger-scale training.
Target Audience
Primary users are researchers and decision-makers in scientific discovery, capital allocation, and policy domains who need counterfactual reasoning and intervention mapping to support judgment under uncertainty.
Features
- Neuro-symbolic architecture combining SmolLM3-3B neural prediction with symbolic reasoning and knowledge graphs
- Proprietary training corpus of ~100,000 historically grounded branching decision scenarios with reasoning graphs and counterfactual options
- Constrains implausible transitions by applying explicit rules over represented authority relations and institutional structures
- Simulates parallel worlds to show how alternative actions change the social environment and available next decisions
- Evaluated on transfer to held-out cases, including option ranking, next-state prediction, and consistency across linked decisions
- Reproducible checkpoints, graph schemas, and training/evaluation pipelines built for new decision domains and larger-scale training