<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Email on Linux Colorado</title><link>https://www.linuxcolorado.com/tags/email/</link><description>Recent content in Email on Linux Colorado</description><generator>Hugo</generator><language>en-US</language><lastBuildDate>Thu, 05 Nov 2026 00:00:00 +0000</lastBuildDate><atom:link href="https://www.linuxcolorado.com/tags/email/index.xml" rel="self" type="application/rss+xml"/><item><title>Forty Cents of Agent Research, One Hallucinated Email, and the Script That Caught It</title><link>https://www.linuxcolorado.com/post/2026-11-05-agent-research-pipeline-deterministic-verifier/</link><pubDate>Thu, 05 Nov 2026 00:00:00 +0000</pubDate><guid>https://www.linuxcolorado.com/post/2026-11-05-agent-research-pipeline-deterministic-verifier/</guid><description>&lt;p&gt;&lt;a href="https://github.com/vlouvet/NeonBench"&gt;NeonBench&lt;/a&gt; is in public beta, and the people most likely to care are the schools and studios that still teach glass bending. There is no directory of them, and finding them is tedious web research, which is exactly what a cheap agent is for, and exactly where a cheap agent will make things up. So I spent forty cents on ten agent jobs, got back a spreadsheet of trade schools, and built a deterministic verifier that refused to trust any of them. It caught a hallucinated email address. The whole thing is the subject of this post.&lt;/p&gt;
&lt;p&gt;The agent is the same Hermes/DeepSeek setup I use for &lt;a href="https://www.linuxcolorado.com/post/2026-11-06-cheap-coding-agent-claude-reviewer/"&gt;code jobs&lt;/a&gt; on the homelab GPU box (&lt;code&gt;.62&lt;/code&gt;): an open-source agent framework in a container, a DeepSeek V4 Flash model through OpenRouter, briefs written as Markdown files, one job per brief, run in parallel. For research jobs it gets the &lt;code&gt;web&lt;/code&gt;, &lt;code&gt;file&lt;/code&gt; and &lt;code&gt;terminal&lt;/code&gt; toolsets and a plain output directory instead of a git worktree. A pleasant surprise: the framework&amp;rsquo;s web search works with no API key at all (it falls back to a free search backend), so the research side cost nothing beyond model tokens.&lt;/p&gt;
&lt;p&gt;The punchline, spoiled up front: &lt;strong&gt;the model followed every rule except the one that mattered, a dumb script was the only thing that noticed, and the forty cents bought a lesson about rules versus controls that I&amp;rsquo;d happily have paid more for.&lt;/strong&gt;&lt;/p&gt;</description></item></channel></rss>