---
license: apache-2.0
task_categories:
  - text-classification
language:
  - en
tags:
  - decision-model
  - calibration
  - onnx
  - laya
pretty_name: laya-onnx decision eval
---

# laya-onnx decision eval

A small, reproducible self-check for [laya-onnx](https://github.com/Geoking2104/laya-onnx) —
an ONNX Runtime backend for **Laya typed System-1 decisions** (`choice` / `score` /
`noul`, no generated tokens).

## Files

| file | what |
| --- | --- |
| `decisions.jsonl` | 8 labeled states / 10 questions with gold answers |
| `results.json` | raw published results (accuracy, MAE, ECE, latency) |
| `results.md` | method + published numbers |
| `checksums.json` | SHA-256 manifest of the `receptron/laya-onnx` bundle |

## Published result

CPU run (Windows 11, Intel i7-1255U, `onnxruntime`), bundle `receptron/laya-onnx`
@ `68f27dfe`:

| question type | n | accuracy | MAE | ECE |
| --- | ---: | ---: | ---: | ---: |
| choice | 4 | 1.00 | – | 0.456 |
| noul | 4 | 1.00 | – | 0.129 |
| score | 2 | 0.50 | 0.530 | 0.112 |
| **overall** | **10** | **0.90** | – | **0.256** |

Latency per `predict()` call: mean 2808 ms / p50 1760 ms / p95 7418 ms.

This is a reproducibility self-check, **not** a leaderboard — the sample is tiny and
the intervals are wide.

## Use

```bash
git clone https://github.com/Geoking2104/laya-onnx && cd laya-onnx
python benchmarks/evaluate.py receptron/laya-onnx --eval decisions.jsonl
```

## License

Apache-2.0.