Introducing Lattice
A question-answering engine without a neural network: knowledge as readable facts, every answer explained.
Lattice is Convergent's core system, and it works fundamentally differently from today's AI models. There is no training, no weights and no language model in the background.
Knowledge you can read
Everything Lattice knows is stored as individual facts, such as "a dog is an animal" or "penguins cannot fly". For each question, Lattice fetches only the facts it needs and derives the answer from them.
Every answer explained
Because the answer is derived from facts, every step can be shown: which facts were used, which rule applied and why a more general statement lost.
More knowledge, same effort
Lattice keeps only the parts of its knowledge in memory that a question needs right now. In our measurements, this working set stayed practically the same size with 351 times more knowledge.
What comes next
The standard model Omni-Compact 1.5 answers 96.2% of our test catalogue. The earlier models Ares 1.7 and Mach 9 Master are retired.
Links