The core ideas
How Masar approaches typed, verifiable generation.
From intent to a verified program
The LLM is System 1 — fast and intuitive. Masar is the System 2 layer beside it: a symbolic compiler and two independent verifiers that build the program and prove it correct. The model only ever proposes typed choices; the guarantees come from the system around it.
Local 1.5B model — no frontier LLM in the loop
Correctness from the compiler + dual verifiers, not the weights
Bounded action space — only declared, verified behaviors
How It Works
Intent
A natural-language request
Narrow
Embedding routing selects candidate behaviors from the typed library
Propose
A small model picks the behavior and its typed parameters — never free-form code
Compile & verify
The deterministic compiler resolves it; two independent verifiers must pass
Run
The generated system runs the closed Event → Guard → Transition → Effects circuit
Built For
Bounded agents
An agent that can only invoke declared, pre-verified behaviors — so its action space is bounded and every output is checked, not merely likely.
Orb construction
AI-assisted .orb construction from a typed behavior library, where the compiler — not the model — guarantees structural correctness.
Local, sovereign builds
Because correctness comes from the IR and verifiers, the model can be a self-hosted 1.5B adapter — private, offline-capable, no frontier cloud model.