A decision model that answers with probabilities, not prose

What it does

@receptron/laya wraps the Laya decision model — a ModernBERT encoder with a decision head — for Node.js and TypeScript. You feed it a state (a support ticket, an email, a JSON object) plus typed questions, and it returns answers with calibrated probabilities in one forward pass: choice picks an option with per-option probabilities, score places the state on an ordered rubric, and noul gives a calibrated P(true) for a yes/no statement. No text generation, no Python at runtime — ONNX Runtime does the inference, and outputs match the Python reference implementation to four decimal places.

The interesting bit

The boring part is the value: it’s a classifier you can interrogate with many questions at once, batched into a single model run (three questions in ~140 ms on Apple silicon, warm). The answer_confidence field is the same quantity — max probability — across all question types, so you can build one confidence-gating rule for everything instead of per-type thresholds.

Key highlights

• Three answer types (choice, score, noul), each with calibrated probabilities and a uniform answer_confidence.
• Runs fully offline after the ~1.7 GB weights are cached; no PyTorch or Python needed.
• Typed responses — a choice question yields a ChoiceAnswer without casting.
• Reproducible loads: pin a revision, point at a local ONNX export, or swap checkpoint variants.
• Ships an ONNX export script (export/export_onnx.py) with a logit-diff check against PyTorch (~1e-5).

Caveats

• Heavy footprint: ~1.7 GB weights, ~2 GB RAM for the loaded model plus a few hundred MB per batch.
• Hard limits: 192 tokens per question’s options, 512-token state truncation, and roughly 20 options max per choice per the model’s own recommendation.
• Confidence values for choice questions with 11+ options intentionally differ from upstream due to temperature clamping (0.945 vs. the checkpoint’s 0.1006, which Python clamps to 0.5).

Verdict

Worth a look if you’re routing tickets, triaging emails, or gating LLM pipelines on calibrated confidence in a JS stack. Skip it if you need generated text or want a model under a gigabyte.