The inventions behind Edwin

Every fact gets two scores. The first measures how well it was analyzed: two or more AI models assess it and their confidence is combined. The second is a separate fact-check that asks a simpler question: is this likely true? Multiplying the two gives an effective confidence. Facts that fail the fact-check (below 0.2) are blocked, and uncertain ones (0.2 to 0.4) are capped at 0.4. The result decides what ranks first, what enters the trusted context the AI reads first, and what gets archived. The filing's example: a well-analyzed false claim (0.85 × 0.2) lands at 0.17 and stays out of the trusted context.

Instead of a single confidence number, each fact is sorted into one of three zones: coherent (settled), contested (disputed or weakly supported) or frontier (too new to judge). A fixed rule order decides the zone from four signals: how much the AI models disagree about the fact, whether it is linked to a contradicting fact, its fact-check score, and how many times it has been reviewed. The zones then guide daily briefings, what gets reviewed next and the labels people see, so “this is contested” and “this is settled” can be said in plain words.

When a new fact arrives, the system gathers the existing facts most likely to conflict with it: facts already in dispute on the topic, confident facts in the same category, and its direct neighbors in the knowledge graph. They go into the fact-check alongside the new fact, so the model judges it against what is already known rather than in isolation.

“Alice manages Project X,” “Alice isn't in management” and “Project X needs a management sponsor” look fine two at a time, but not together. The system groups three to five closely linked facts and asks a language model to check them as a set, then labels anything it finds as temporal, numerical, semantic or cross-referential. One check per group can replace or supplement comparing every pair.

Each fact gets a priority from how often it is used, how long since it was reviewed, how connected it is and whether it is flagged for a contradiction. That priority is tripled when the AI models sharply disagree about the fact and halved when they firmly agree. Every batch also includes at least one brand-new fact, so new knowledge is never crowded out by established facts.

Every fact carries two dates: when it was true in the world and when the system learned it. When two facts seem to clash, both dates go to the model, which can answer contradiction, update, consistent or unrelated. An update links the newer fact to the one it replaces, and the older fact stays on record with its date. A budget that went from $500K to $750K is history, not a conflict.

Claims 1 to 6 and 9, joined into one cycle. The knowledge graph shapes verification, and verification reshapes the graph. Disagreement between the models decides what gets reviewed next, and review uncovers new links. Dates keep facts current, zones track the health of the whole, and scheduled challenges send weak facts back for another look. Each part uses the others' outputs.

A snapshot of the knowledge base is stripped of personal details such as names, emails and phone numbers, encrypted under a key the user controls (AES-256-GCM), signed to prove where it came from (Ed25519), and stored on decentralized storage (the Walrus protocol on Sui). To restore it, the user downloads the snapshot, verifies the signature, decrypts it and rebuilds the knowledge base.

Like the brain consolidating memories during sleep, the system regularly revisits facts it has come to rely on (the filed default is every six hours). A language model is asked to argue against each one, with its related facts in view, and to score how vulnerable it is. Facts that don't hold up are demoted, facts that do are kept, and every decision is logged with the model, the score and the reasoning.