The TOPS Trap: Local AI Models Hallucinated Facts, Cited Real Sources, and Sounded Confident
Washingtoner/10338033

SPOKANE VALLEY, Wash. - Washingtoner -- AIToolworx today published The TOPS Trap, a reproducible study of whether language models running on local hardware stay inside the evidence they are given. In the configurations tested, they often did not, and their answers read as though they had.

The study gave six local model configurations on a Raspberry Pi 5, some on the CPU and some on the Hailo-10H accelerator, a fixed, hashed folder of evidence and one rule: answer only from it.

"Our assertion is not that AI sometimes gets facts wrong. It is that an AI system can produce a well-formulated, seemingly well-reasoned, and convincingly supported answer while contradicting or misrepresenting the very evidence it was explicitly instructed to follow." — James Childers, Founder, AIToolworx

More on Washingtoner

• Author Robert A. Robinson Introduces Anarchy at the AYP:
• Renew Integrative Health Publishes Updated Peptide Therapy Guide to Help Patients Understand Their Options
• AI Agents Are Getting Their Own Computers. HRTech Leading in Vertical Harness Development
• UK Financial Ltd Announces Two-for-One Upgrade of Maya 3 to Maya 3G Token Using Chainlink Cross-Chain Technology
• Critical FDA Milestone much Closer as Preservative-Free Ketamine Review Advances: NRx Pharmaceuticals (N A S D A Q: NRXP)

What the tests found:

• Who is actually building your local AI model? The retail box reads Raspberry Pi, Empowered by Hailo, Made in Taiwan, with no mention of Alibaba or China. The Qwen models tested on that board were built by Alibaba, and the instruction sentence found in the prompt names Alibaba. Where hardware is made is not the supply chain. The study does not show who inserted the sentence or why.
• Citations were not proof. In one evidence-packet test, 45 of 93 claims (48.4%) were unsupported or contradicted, and 18 of 37 cited claims (48.6%) pointed to a supplied source that did not support them.
• In separate closed-book tests, one model stated in all four runs that Andrew Jackson was impeached, with no hedging.
• In 32 of 32 runs of two Qwen 2.5 configurations started with no system message, an identity sentence the user never typed was added to the prompt.
• Speed did not predict trust.

Findings apply to the exact configurations tested, on one platform and two topic packets, not to all local AI. The study is not peer reviewed. AIToolworx has a commercial interest in AI security consulting. Before publication, the author offered Raspberry Pi, Hailo, Alibaba/Qwen, DeepSeek, Meta and Mistral AI the chance to comment; none responded.

More on Washingtoner

• Spokane City Council Passes Community Safety Sales Tax
• AutoProIT Achieves Cyber Essentials Certification, Reinforcing Its Commitment to Secure Digital Transformation
• DayPass llega a ChatGPT y Claude
• DayPass launches in ChatGPT and Claude
• Renkeer Water Quality Sensors Support Efficient Operation of AI Liquid Cooling Systems

The paper, a 12-page executive summary and the evidence archive are free at http://tops-trap.aitoolworx.com/downloads/ with a SHA-256 hash beside each file. Permanent copy: https://doi.org/10.5281/zenodo.23423297

About AIToolworx: AI architecture, data security, and private and edge computing. Secure AI. Local control. Trusted data.

Cover illustration and diagrams were AI-generated. Company and model names are trademarks of their owners; no endorsement implied.