Refleqt Labs is a small research programme with one premise: the transformer became a language artefact by accident, and the more interesting object is the one underneath. A model that thinks in algebra rather than in tokens. A model that knows what it knows. A model whose every assertion is traceable to an experiment, a seed, and a stage gate. We are not optimising a chatbot. We are characterising a substrate.
A knowledge transformer is a stack of small, audited components — one per algebraic domain. Symmetric groups S_n. Knowledge-graph triples. Molecular SMILES. Text-to-SQL. Description logic. Each domain has its own data generator, its own attention-mass profile, and its own LoRA adapter on a shared base model.
We have trained five of them on Mistral-7B; AQL decoding accuracy is above 95% on every one (D1: 100%, D2: 97%, D3: 98%, D4: 100%, D5: 100%). The ambition is not to fit a single corpus. It is to compose the pieces — to ask whether a model that has a group-theory adapter and a graph adapter can solve a problem neither one of them could solve alone.
Every empirical claim flows through a four-stage gate. Characterise a domain. Verify the circuits that implement it. Establish sheaf consistency across domains. Demonstrate composition. A claim does not advance one stage until it has cleared the one below.
We treat contradictions as discoveries. Block-diagonal FFNs win 8 of 17 head-to-head downstream tasks at 125M and collapse to chance on S₄ group composition; the apparent contradiction is in fact the most informative single result we have, because it tells us where the inductive bias starts and stops working. We publish the underpowered comparisons, the failed symbolic recoveries, and the architectural collapses with the same prominence as the wins.
We will not annotate by hand. We will not pay for closed APIs. We will not run two experiments on a single GPU at the same time. We will not publish a number without the seed, the model id, and the runtime that produced it.
The interesting frontier is not larger language models. It is verifiable knowledge models. A system that can be queried about a finite group and respond with the correct multiplication table — and can be asked, separately, whether it knows the answer — is qualitatively different from a system that interpolates a plausible string. We are interested in the qualitative difference.
If the programme works, the artefact is not a product. It is a class of architectures, a stack of trained domains, and a record of the experiments that justify each one. If the programme does not work, the record will say so — in the same vocabulary, on the same site, with the same prominence.