Open the first-use page · 中文入口
Run two synthetic loyalty tasks offline, then inspect the outcome and evidence. Start with the two tasks and commands, inspect fixed starting data and fields, then read the report. The next-scenario sketch is a proposal only, not implemented and not additional benchmark coverage.
Download and extract the 0.5.0 source ZIP first. From the extracted archive root, Python 3.9 or later is sufficient; no dependencies, model service, credential or network is required. A fresh environment is optional:
python3 -m venv --without-pip .venv
.venv/bin/python scripts/export_fixed_data.py --check
.venv/bin/python scripts/run_demo.py --task draft_for_approval --agent reference --task-language en --output results/draft-en
.venv/bin/python scripts/run_demo.py --task execute_approved --agent reference --task-language en --output results/execute-en
For the intentional failure, use --agent always_refuse on the approved task
and a new output directory. Exit code 1 is expected. Each run writes matching
JSON and Markdown reports. CLI language defaults to Chinese; use
--task-language en for English. --report-language defaults to task language.
The first-use page defaults to English and provides a Chinese switch.
This version contains one business template, two approval-state tasks and three datasets
(baseline, partial_refund, contact_kinds): six task/data combinations. Select
--dataset partial_refund or --dataset contact_kinds; omitted means baseline.
The corresponding fault scripts are ignore_partial_refund and service_as_marketing.
Their baseline passes are not evidence of correct variant handling.
See the first-use exercise, feedback,
integration and version notes.
Code 0.5.0 / task prototype-v0.1 / language task-language-v0.4.0 / data fixed-v0.2.
Historical model evidence used code 0.3.0 and is retained in the separate historical preview; it has not been relabeled as 0.5.0.
Current script examples and historical model results remain separate. No real model is called.
Script success does not establish model quality or production safety; synthetic
business rules still require expert review.
Full source and fixed exports are for maintainers. Normal agents receive only
the public specification under agent-input/ and controlled tool results, never
reference answers, scoring internals or report examples. This is not a sandbox.
See build/integrity notes, license scope, Apache-2.0 and CC BY 4.0.