RustXNG + Gemini Decision 1 cuts reading by 86% in test

The experiment addresses a practical bottleneck: when an AI agent researches a topic, reading every search result fills its context quickly. RustXNG, a self-hosted meta-search engine released on Oct 18, 2026, returns 35 to 125 results per query in 1.3 to 2.9 seconds, which is fast but creates more candidates than a large model should read. The writer's script uses Gemini Decision 1 to assign relevance scores and injection-likelihood checks to each candidate, plus a single choice question for the best pick. In a test with five normal snippets and one injected with a bias instruction, the injected candidate was excluded with an injection suspicion of 0.969, while all other candidates scored 0.05 or below. In a freshness-sensitive query about new AI models in Nov 2026, the harvested results were mostly October roundups and old announcements, and the model returned 'other' as the best pick with 0.87 confidence rather than forcing a wrong answer. The writer notes the official best practice recommends handling date and number comparisons in code and passing results as input, and plans to add publication dates to the input next. The comparison with Laya, a local judgment-only model, shows a trade-off: Laya is fully local and free for private text, but in this test it compressed relevance scores and flagged all six candidates as injection suspects, possibly because the question wording was too aggressive. The writer's conclusion is to use Decision 1 for large-scale sorting of public web text and Laya for surface classification of private text.