cl33-opLM: a model you can read

A transformer computes as usual, but it is not allowed to speak. At every position it writes 32 operators, each 15 coefficients in the Cl(3,3) geometric algebra, and the part of the model that produces words sees only those operators. Zero them and the model has no path to its output. That claim can be tested on the public release.

The current base has 470 million parameters in 29 layers, trained to 20 billion tokens. Its chat version, P1, is fine-tuned with the part that reads the operators frozen, so everything the model has ever written stays decodable.

The constraint has a price. At 50 million parameters after 3 billion tokens, the operator model's validation loss is 3.3 % above a matched ordinary transformer (one seed), and the gap shrank steadily as training went on. One thing is not shown: that this particular algebra does anything a random basis of the same size would not. Four matched comparisons tied or lost. The bottleneck is established; the algebra's advantage is not.

A fact is stored as the model's own operators over it, and recalled by replaying them into the same readout: no extra network, nothing trained. Lookup is a fixed rule with thresholds and the option to abstain. Every write, read and deletion leaves a receipt, and deleting a memory leaves the model exactly as if it had never been stored.

On top of that, the frozen model learns multi-step procedures from unlabelled experience (a search checked by a verifier, a statistical test before anything is kept) and stores them as objects that can be called, inspected and deleted like facts.

These results come from synthetic test worlds, mostly on one seed. In free conversation the memory almost never acts: 0 of 341 natural questions triggered it, and 0 of 47 natural corrections were captured. The chat version is not finished, so it is not released: asked about things it had never been told, it still answers anyway (0 of 8 in an internal review). What to store from real conversation, and how a model earns the authority to say "I wasn't told that", are the open problems.

The newest step asks whether images can enter the same causal object, rather than a separate one bolted beside it. A frozen image encoder (DINOv2) sits outside the model, the way a tokenizer does for text. A small projector of 3.7 million parameters, trained only through the frozen chat model, turns the encoder's features into 17 input tokens. P1 then writes its own memory of the picture with the same capture it uses for text, and the unchanged memory machinery stores it, links it to text facts and deletes it exactly.

With the right image, P1's captions become far more likely (2.43 nats, against 4.08 for a wrong image and 4.00 for none). A memory P1 writes of a photo chains through a stored text fact about the object in it 96.2 % of the time, against 21.3 % with the wrong photo, matching a separately trained vision module (95.9 %), which was then retired. It took four preregistered stages to get there: 45.1 %, then 79.7 %, then 95.9 % at a captioning cost that failed its guard, then 96.2 % inside it.

The limits: a closed vocabulary of 79 object kinds, two-step chains and one image set (COCO). New kinds are only partly recognised: no zebras (0 of 24) and no giraffes (0 of 40). The frozen encoder does the seeing, so seeing is not claimed for cl33. Images are not in the served model yet.

Whether the algebra itself matters, measured against a random basis of the same size.

What to remember from natural conversation, and when memory should act.

A reliable way for the model to say "I wasn't told that".

Whether the memory and procedure results survive a second world, a second seed and natural text.

Whether a frozen image generator can serve as a second readout of the same memories (proposed, not built).