Skip to main content
The laya driver connects Polyglot Decision to a separately deployed typed-decision-lab service. The Python service is maintained outside the InstructorPHP repository and Composer packages. One PHP driver works with either the Laya-MLX or upstream PyTorch adapter because model selection, loading, hardware, queues, and timeouts belong to the external service. PHP never starts Python, invokes uv, downloads weights, selects a device, or falls back to another provider.

​
Run the external service

Obtain the separately distributed typed-decision-lab repository from your deployment source. It has its own version, Git history, pyproject.toml, uv.lock, CLI, tests, and runtime extras. From that repository, configure one pinned adapter, route, and checkpoint revision:
Run uv run td-lab serve. Use that repository’s README for installation, authentication, resource limits, adapter extras, and real-checkpoint smoke commands. /healthz reports process liveness, while /readyz becomes successful only after the checkpoint is loaded.

​
Configure PHP

Point the bundled preset at the already running service:
The default preset requests the explicit laya-typed-decisions route. Other bundled model records are laya and laya-multilingual. The service accepts only its configured route; there is no automatic language or workflow routing. Every question needs instructions because both reviewed Laya implementations require that field. Text, object, and list state retain their System One shape rather than being converted to classifier text.

​
Model action probability

Laya returns a separate action-head probability for each answer. Polyglot keeps it as a typed auxiliary signal:
This value does not replace a Choice, threshold a Noul, change a Score, or grant permission to perform an application action. Caller-owned policy decides if and how it is used. The primary answer and its distribution remain authoritative. The adapter validates the only documented extension, action.act_probability, and checks Laya’s redundant Noul confidence against the confidence derived from its probability. Unknown fields, incomplete answers, and inconsistent values fail closed.

​
Operational attribution

The returned model identity is route@checkpoint-revision, not the ambiguous upstream string laya-rl-agent. x-request-id becomes the provider request ID. Backend, device, precision, revision, queue, inference, and server timing facts remain available on DecisionResponse::responseData() and the external service’s /v1/models and /stats endpoints. Record those exact deployment facts when using the provider evaluation workflow. Contract parity between MLX and PyTorch does not establish equal calibration, latency, or application quality.