<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Raed Alhamdan</title>
    <description>The latest articles on DEV Community by Raed Alhamdan (@raed_alhamdan_be31f9ae44c).</description>
    <link>https://dev.to/raed_alhamdan_be31f9ae44c</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4213338%2Fe0e260b4-2750-4510-bff6-b69aed76ef7b.png</url>
      <title>DEV Community: Raed Alhamdan</title>
      <link>https://dev.to/raed_alhamdan_be31f9ae44c</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/raed_alhamdan_be31f9ae44c"/>
    <language>en</language>
    <item>
      <title>PHANTOM TRACE: Finding the Ghost Between an Agent's Claim and Its Evidence</title>
      <dc:creator>Raed Alhamdan</dc:creator>
      <pubDate>Wed, 28 Oct 2026 15:58:22 +0000</pubDate>
      <link>https://dev.to/raed_alhamdan_be31f9ae44c/phantom-trace-finding-the-ghost-between-an-agents-claim-and-its-evidence-118o</link>
      <guid>https://dev.to/raed_alhamdan_be31f9ae44c/phantom-trace-finding-the-ghost-between-an-agents-claim-and-its-evidence-118o</guid>
      <description>&lt;p&gt;&lt;em&gt;This is a submission for the &lt;a href="https://dev.to/challenges/hacktoberfest-week4-2026-10-26"&gt;Hacktoberfest Open-Source AI Challenge Week 4: Ghost in the Machine&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Update (Oct 29):&lt;/strong&gt; the local AI review now uses an open NLI model and has been verified end-to-end in the browser. The earlier version could not produce a usable model response.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I Built
&lt;/h2&gt;

&lt;p&gt;PHANTOM TRACE reviews AI-agent execution traces and checks whether an agent's claims are supported by the events they cite. It flags structural problems such as missing references, and offers an optional open-model review for claims and cited evidence. The “ghost” is the gap between what an agent says happened and what the trace actually records.&lt;/p&gt;

&lt;h2&gt;
  
  
  Demo
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://ralhamdan1412-eng.github.io/phantom-trace-challenge/" rel="noopener noreferrer"&gt;Try the live demo&lt;/a&gt;. It opens with a synthetic trace containing a missing-reference error, so you can see the deterministic audit immediately. Select &lt;strong&gt;Run local AI review&lt;/strong&gt; to run the open model in your browser. On the demo trace it marks EV-002 as contradicted by EV-001 (the deployment check failed), EV-003 as insufficient (its cited evidence EV-404 does not exist), and EV-005 as supported by EV-004. The first run downloads about 244 MB of model files.&lt;/p&gt;

&lt;h2&gt;
  
  
  Code
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://github.com/ralhamdan1412-eng/phantom-trace-challenge" rel="noopener noreferrer"&gt;Source code on GitHub&lt;/a&gt;. The separate public repository contains the application, public documentation, tests, dependencies, and MIT license.&lt;/p&gt;

&lt;h2&gt;
  
  
  How I Built It
&lt;/h2&gt;

&lt;p&gt;The app uses React and TypeScript. Deterministic checks validate trace structure and evidence references. The optional AI path uses Transformers.js with an open natural-language-inference model, &lt;a href="https://huggingface.co/Xenova/nli-deberta-v3-base" rel="noopener noreferrer"&gt;&lt;code&gt;Xenova/nli-deberta-v3-base&lt;/code&gt;&lt;/a&gt;. It is an ONNX conversion of the Apache-2.0 &lt;a href="https://huggingface.co/cross-encoder/nli-deberta-v3-base" rel="noopener noreferrer"&gt;&lt;code&gt;cross-encoder/nli-deberta-v3-base&lt;/code&gt;&lt;/a&gt; and runs entirely in the browser. Each claim is scored against every existing evidence event it cites. Entailment, contradiction and neutral map to supported, contradicted and insufficient. A verdict below 80% confidence is reported as insufficient. Every result then passes a deterministic validator that rejects unknown, uncited or missing evidence IDs. If one claim fails, it is marked unassessed without affecting the others.&lt;/p&gt;

&lt;p&gt;I first tried a small generative model, SmolLM2-135M-Instruct. In real runs it never produced the required JSON. When limited to the three labels, it chose "supported" even for clear contradictions, so I replaced it with a model built for this exact task. The model files come from Hugging Face and the ONNX Runtime files from jsDelivr on first use. Trace text is never sent to an inference API. The model can still be wrong: it matched 8 of 11 labels in my synthetic check. Its results are hypotheses for a human to verify.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why Does Open Innovation Matter?
&lt;/h2&gt;

&lt;p&gt;Open weights make the review path inspectable and let people change or replace the model. Keeping inference in the browser is designed to avoid sending trace contents to a hosted AI service; only the model files (from Hugging Face) and the ONNX Runtime WebAssembly files (from jsDelivr) are fetched on first use. PHANTOM TRACE separates clear structural errors from uncertain model assessments, which remain hypotheses that need a human to verify against source evidence.&lt;/p&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
    </item>
  </channel>
</rss>
