Laya — Lab Component
Lab component — not in the Synedrix OS tech stack·Evaluated under LAB-003·Not required in any deployment mode
Not a Synedrix OS platform. Laya is not one of the twelve platforms and holds no OS layer. It is evaluated under
LAB-003alongside Jev, in structural isolation from production, as a candidate System One decision model. This page is lab reference material, not a stack platform page, and not citable as authority by any canonical document.
What Laya is
Laya (laya on PyPI, from Convai Innovations) is a System One decision model: a non-autoregressive model that
answers typed questions about a piece of state in a single forward pass and returns a probability with every
answer. It generates no text. The package is published under Apache-2.0 and its checkpoints are open weights that
can be run on your own hardware.
| Concept | What it is |
|---|---|
| Typed questions | Three kinds: choice (pick one label from a set you define), score (rate against an ordered rubric) and noul (a yes or no probability) |
| Checkpoints | Three, per the project’s own documentation: laya (English, ModernBERT-large, about 421M parameters), laya-multilingual (mmBERT-base, about 322M, long inputs) and laya-typed-decisions (fine-tuned for typed-decision workflows) |
Router | Picks a checkpoint per request, for example by detected language |
laya-serve | An optional HTTP server. Its authors built it to speak the same request protocol as TypeSafe’s Jev, so a client written for Jev can point at a self-hosted Laya server. It supports an optional bearer key and a strict mode that drops Laya’s extra fields |
| Confidence | On choice and score answers, Laya’s confidence is normalised entropy, not Jev’s formula (laya-serve adds x_jev_confidence with Jev’s formula for the same probabilities), so a threshold tuned on one model does not transfer to the other |
| Fine-tuning | The project ships notebooks and scripts to fine-tune a checkpoint on your own labelled decisions |
What Laya is normally used for
- Routing and triage — sending a ticket, message or task to the right queue or agent
- Classification and scoring — labels and rubric scores where a calibrated probability matters more than an explanation
- Gating — yes or no checks before a more expensive step runs
The common thread: bounded decisions made many times, where a probability the calling code can compare with a threshold is more useful than a paragraph of generated text.
How Synedrix OS uses it
Synedrix OS does not use Laya in production. Under LAB-003 it runs as a self-hosted server inside the lab’s
isolated environment, on synthetic data, and is called only by unregistered lab agents.
What it owns
Nothing in production. Inside the lab, it answers routing, screening, verification and escalation questions for lab agents, so its accuracy, calibration and latency can be measured against Jev and against LLM-only baselines.
What it does not own
| Concern | Owner |
|---|---|
| Whether any action is permitted | Policy Guard Agent (os-agt-002) and HITL Gateway Agent (os-agt-004). A decision model’s confidence is evidence, never permission |
| Committing durable memory | Memory Curator Agent (os-agt-010) |
| Model routing policy for registered agents | The agent’s runtime policy (Agent Runtime Policy Spec) |
No registered agent may hold Laya, or a tool wrapping it, in its tools_allowed.
Why Laya is evaluated
- It can run in every deployment mode. Laya is self-hostable, including air-gapped, where no outbound model call is allowed. That fits Principle 11, “Open by default, sovereign by design”, in a way a managed API cannot.
- Quality out of the box is the open question. On its authors’ own published benchmark, the base checkpoint scores well below Jev until it is fine-tuned on decisions like the ones it will make.
LAB-003measures fine-tuned Laya against Jev on its own held-out data. - Adoption, if it ever happens, is a separate decision. A positive result from the lab would be an input to a normal ADR. The lab cannot adopt anything by itself.
Licence
The laya package is Apache-2.0, read from its package metadata. The licence of each checkpoint on Hugging Face has not been verified and must be read from its model card before any fine-tuned checkpoint is kept.
Official documentation
layaon PyPI — package and licence metadata (Apache-2.0)- Laya documentation
- Laya on Hugging Face — checkpoints and model cards
Synedrix OS · Lab component reference · Laya