Proofline Qwen Autopilot

Inspiration

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for Proofline Qwen Autopilot

Inspiration

News headlines compress complex public decisions into a sentence. Citizens, journalists, researchers, and policy teams still have to search scattered Diet records, identify the right bill or policy topic, read long proceedings, and determine what the evidence actually supports. Proofline Qwen Autopilot turns that manual workflow into a verifiable research case.

What it does

Enter a news URL, policy or bill name, or Japanese Diet member name. Proofline:

1. identifies the formal policy topic instead of trusting only the headline;
2. retrieves related official National Diet records;
3. sends a compact evidence packet to Qwen 3.7 Plus;
4. validates every evidence ID returned by the model;
5. separates verified findings from unresolved questions;
6. creates a timeline, speaker statistics, legislative trace, and source-linked report;
7. keeps a human approval checkpoint before the result is treated as final.

A misleading or incomplete article cannot become the sole source of truth: Proofline checks the claim against primary records and explicitly says when the retrieved evidence is insufficient.

How we built it

The frontend is React + TypeScript. Search and evidence normalization use the National Diet Library API. Qwen 3.7 Plus runs through Alibaba Cloud Model Studio, with the production analysis endpoint deployed on Alibaba Cloud Function Compute.

Architecture:

News URL / policy / member → topic extraction → Diet record retrieval → evidence packet → Qwen 3.7 Plus → evidence-ID validator → human approval → source-linked report

The public application reports the active provider and model, and the repository includes deployment code plus an architecture diagram.

Technical depth

• Evidence-constrained JSON output with low-temperature generation
• Server-side Qwen credentials and a production Function Compute proxy
• Retrieval/result-count separation so “matched” never means “verified”
• Evidence-ID validation that rejects unsupported model references
• Graceful fallback when records or AI output are incomplete
• Japanese public-record parsing, speaker aggregation, timelines, and legislative tracing

Challenges

Japanese policy language differs between headlines, bill names, committee debate, and colloquial descriptions. We built staged query expansion and a formal-name resolver so the agent does not simply repeat the article title. We also reduced the evidence packet and increased the structured-output budget to avoid truncated JSON while keeping inference efficient.

Accomplishments

Proofline is a working public application using live Qwen analysis, real Diet records, and Alibaba Cloud deployment. Every major conclusion remains traceable to an official statement, while uncertainty is shown instead of hidden.

What we learned

For public-information research, a confident answer is less valuable than a verifiable one. Production AI needs provenance, uncertainty boundaries, and human approval—not just fluent summaries.

What's next

Expand to ministries and local assemblies, add multilingual research views, monitor newly published records, and compare media claims with official evidence over time.

Live app: https://proofline-qwen-autopilot.hello-you-smap.chatgpt.site
Source: https://github.com/gameyousmap-hash/proofline-qwen-autopilot

Built With

• alibaba-cloud
• function-compute
• next.js
• qwen3.7-plus
• react
• typescript