PHANTOM TRACE: Finding the Ghost Between an Agent's Claim and Its Evidence

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 4: Ghost in the Machine.

What I Built

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.

Demo

Try the live demo. It opens with a synthetic trace containing a missing-reference error, so you can see the deterministic audit immediately.

Code

Source code on GitHub. The separate public repository contains the application, public documentation, tests, dependencies, and MIT license.

How I Built It

The app uses React and TypeScript. Deterministic checks validate trace structure and evidence references. The optional AI path uses Transformers.js with HuggingFaceTB/SmolLM2-135M-Instruct, an open-weight model with an Apache-2.0 license, in a browser-local Q4 configuration. Model files are downloaded from Hugging Face the first time; trace text is not sent to an inference API. Tests cover bounded prompts and reject unknown or uncited evidence IDs. In this deployment check, the first model load did not complete, so I have not verified an end-to-end model response; the deterministic demo is available and working.

Why Does Open Innovation Matter?

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 are fetched from Hugging Face 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.