Status & Roadmap
Masar is research. This page is the honest accounting of what actually works today versus what is still being figured out. We would rather under-claim here than publish numbers we cannot stand behind.
Component status
| Component | Status | Reality |
|---|---|---|
| Knob-value picker (small fine-tuned model, per-knob selection) | Production-ready (narrow) | Trained, deployed, and clears its gate on service-pattern behaviors. Not yet general across app behaviors — that is the open work. |
| Embedding behavior router | Production | Picks the right behavior by cosine similarity of intent vs. behavior descriptions. No training, no drift. |
| Orb compiler + dual verifiers | Production | The deterministic compiler and the compile-time + runtime verifiers are the backbone correctness comes from. |
| Tool-calling subagent (repair turns) | Trained, dormant | Reproduces the call-and-repair format on held-out data, but the coordinator's deterministic fast path means its inference seam is rarely invoked. End-to-end battery eval pending. |
| GFlowNet construction | Experimental | Implemented and unit-tested, but largely superseded by the factory layer, which guarantees structure by construction. |
| Composition models (linear / MLP / GNN) | Experimental | Strong on planted/synthetic tests; not yet integrated into the production coordinator. |
| JEPA world model | Parked / gated | The PoC did not validate; a redesign made the architecture correct but it is below the self-supervised data floor. See The JEPA bet. |
What is actually callable today
Masar's services are internal research infrastructure, reached by Almadar's own pipeline — not a public, authenticated product. There is no SDK, no API keys, and no hosted multi-tenant API. Treat anything below as internal and subject to change.
The API we would build (roadmap, not shipped)
If and when Masar is exposed as an interface, the right shape is operations on the typed schema itself, exposed in maturity stages — never a generic ML-inference SaaS. Sketch:
- Stage 1 (closest to real):
validate(delegates to the canonical compiler),embed(structural embedding),similar(behavior matching),compile(schema → program). These map to capabilities that exist today. - Stage 2 (designed):
predict-config(knob values),evolve(apply an intent delta),repair,compose. - Stage 3 (research): episodic memory and JEPA-based planning — only after they validate.
Hard prerequisites before anything is public: real authentication and tenancy; durable persistence (today's memory is ephemeral); the corpus-parser gap resolved; and metrics backed by the verifier battery rather than asserted.
Follow along
Masar is built in the open. Watch the work on GitHub, or reach us at hello@almadar.io.