Masar
Masar is an open research program, not a commercial product. There is no hosted API, no SDK, and no sign-up. These pages describe the ideas and their honest current status. See Status & Roadmap.
Masar is the research arm of Almadar: the models and methods that help an agent construct software by composing Orb — a typed, compiler-verified intermediate representation — instead of writing free-form code.
The thesis is neurosymbolic: a small neural model proposes only typed choices, while a symbolic IR, a deterministic compiler, and a dual-verifier loop construct the program and prove it correct. Correctness lives in the representation and the verifiers, not in the model's weights — which is why a fine-tuned 1.5B model, run locally, is enough.
The core ideas
- How it works — the neurosymbolic loop (neural proposer + symbolic compiler/verifier), and the other paradigms we borrow from.
- Execution-grounded fine-tuning — why we decompose a frontier LLM's authoring into small specialized models, and train them on signals that come from a real compiler, not from labels.
- The JEPA bet — a world model that predicts build outcomes before they happen, why it is parked, and the conditions that would unlock it.
- System 2 for agents — Masar as the deliberate, verifiable layer beside a fast LLM.
Honesty first
Most of what Masar explores is research. Some of it works in production today; some is experimental; some is parked pending more data. We say which is which on the Status & Roadmap page rather than advertise numbers we cannot stand behind.