Metis's white paper proposes an explicit "knowledge ledger", analogous to a game board, that records what the system knows, how confident it is and which questions remain open. A separate module checks new entries against calculations, formal proofs or empirical evidence, and when an assumption is dropped, every conclusion resting on it gets re-evaluated. The system chooses its next "move", which might be a question, an experiment or a challenge to an assumption, based on whichever reduces uncertainty most, ending in either an answer or a justified abstention.

The company argues that more data and compute will not fix language models' core problems: fabricated answers, unverifiable reasoning chains and knowledge locked inside model weights. It points to related existing work, including Microsoft's GraphRAG, which organises knowledge in a graph outside model weights, and Google's AlphaGeometry2, which pairs a language model with a symbolic engine that verifies mathematical proofs. Metis wants to extend those principles to open-ended real-world questions.

Important caveats apply. The paper describes only a research programme: there is no working system, no benchmarks and no comparison with current reasoning models. The team itself acknowledges that the real world, unlike Go, offers only partial information, that gathering evidence can be costly, and that some questions have no answer. The open question is how a verified knowledge ledger would work in messy domains, given that earlier symbolic AI systems struggled precisely because all their knowledge had to be entered by hand.

Graepel, who also holds the machine learning chair at University College London, is raising a low tens-of-millions round for Metis, with a larger follow-on reportedly in the works, according to Bloomberg. He is the latest architect of DeepMind's best-known breakthroughs to strike out independently, following David Silver's Ineffable Intelligence and Tim Rocktäschel's Recursive.