← all projects

KV4

A claims engine that reads research notes and proposes claims, but promotes one only when enough separate sources back it. A model can miss things, but it can't invent them.

The problem

Most AI-assisted systems blur knowledge, reasoning, and action together, which makes them hard to govern and impossible to reproduce.

Approach

• Research notes go in as Markdown. A model reads each block and pulls out pains, signals, facts and fixes, each tied to a business outcome and to a word-for-word quote.
• Any unit whose quote isn't actually in the source is dropped. A weaker model costs missed findings, never invented ones.
• The buyer's pains are paired with the seller's own case studies on the same outcome. Pairs that don't share an outcome are refused.
• A claim moves up only as more separate sources back it. There are no made-up confidence scores.
• Every confirmation or contradiction is a new row in an append-only ledger. Nothing is deleted, and a link that was demoted can come back.
• The model runs locally, with a guard that refuses a cloud endpoint.

Build notes

• A working engine with tests, Markdown in and claims out.
• It never clears a claim for external use. Only the Governed Research Engine's claim policy can do that.

What's next

• Feed its claims into the Research Engine's review queue
• Map GROMA onto the same ledger