SIEVE

The problem

Search tools for agents return ten links or a paragraph of generated summary. Both fill the context window, and neither tells the agent which sentence to trust.

What I built

• A four-stage investigation pipeline: Scout, Read and Filter, Cross-examine, Judge and Quote.
• Answers in several shapes: numbers, yes/no, timelines, comparisons, code.
• Quotes are copied verbatim from the source. No generated text sits between the page and the caller.
• About 8 seconds median response time.

How it works

1. 01ScoutFind candidate pages for the question.
2. 02Read and filterOpen each page and drop the unreliable ones.
3. 03Cross-examineCheck whether independent sources agree.
4. 04Judge and quoteReturn one exact quote with evidence, or nothing.