Flagship · In development
Lyra
A physics specialist. Lyra is a language model fine-tuned to be the collaborator you want at the whiteboard. It reads the problem, checks the work, and explains the result. It runs on open-weight bases through QLoRA, trained on arXiv, mathematics, and the scientific literature.
Specialization buys a stricter standard. Lyra is judged the way physicists judge each other, on whether the answer survives checking. The metric is verified-correct rate. The failure it is built to kill is the confidently wrong answer.
- Base
- Qwen3-class open weights
- Method
- QLoRA fine-tune
- Corpus
- arXiv, mathematics, science
- Judged on
- verified-correct rate
- Status
- in development
Research · In build
Eris
A research assistant for general relativity, trained from scratch at the byte level and wired to a symbolic engine that performs the calculations. Eris reads the problem, selects the computation, and interprets what comes back. The numbers arrive from the engine, carrying provenance you can point at.
The channel delivering those results is provably inert. Carrying nothing, its effect on the output is exactly zero in IEEE-754 arithmetic, verified in two independent frameworks on the same checkpoint and enforced on every build. Switch it off and the injected values vanish while ordinary prose stays bit-for-bit identical. That is what makes the numerical output auditable.
The symbolic engine is our own, built here and validation-gated. The architecture is assembled, the training corpus is rebuilt and verified, and the second run is staged.
- Domain
- general relativity
- Built
- from scratch, byte-level
- Parameters
- 210M
- Channel
- zero trainable parameters
- Inertness
- exactly 0.0, two frameworks
- Status
- run two staged