Intelligence Workbench

Intelligence Workbench takes a research question, fetches source documents, extracts a structured graph of entities, claims, evidence and causal links from them, and stores it in a bitemporal corpus that can be queried as of any earlier time. From that graph it renders an 11-section analytical briefing.

The system does not forecast. The briefing renderer is deterministic code, not a language model, and cannot produce a probability by construction; the worker also drops any model-written claim that uses forecasting language. What it surfaces is evidence, causal structure, consensus, dissent and how well each claim is supported, and it leaves the judgment to the reader.

It runs as two processes that share one contract. A Rust service owns the database, the HTTP API and the briefing renderer. A stateless Python worker fetches the documents and extracts claims with a language model; through optional gates it also drops irrelevant or prompt-injected documents and scores each claim's support. It is built to be embedded in a forecasting pipeline; the Betomcat bot is one caller.

Built with: Rust (Axum), Python, the mnestic graph store queried with Datalog, SQLite, AskNews.

Code: github.com/Enhso/iw