1. Ingest the world as atomic claims
Sources are not dropped into a generic document pile. The system extracts small, attributable claims with a subject, relationship, object, observation date, confidence, and original provenance. Contradictory claims can coexist.
2. Connect atoms into falsifiable molecules
A molecule is a synthesis: a causal chain, constraint map, industry structure, or emerging thesis. It must preserve the atoms that support it and expose the missing evidence that would weaken it.
3. Direct attention with founder judgment
Outlier choices, declared biases, decisions, and postmortems form the proprietary layer. They tell the system what is worth looking for. They never become factual evidence and cannot be authored by the model.
4. Generate candidates, not commands
Models can compile a frozen evidence packet into an opportunity candidate or article draft. The output remains inference. It has no authority to publish, trade, alter an allocation, or rewrite history.
5. Resolve and retain
Forecasts, rejected ideas, and capital decisions remain in the ledger after the outcome is known. Over time the system can examine where the evidence was weak, where judgment was early, and where conviction was misplaced.