Full paper (PDF)
- Identifier
- AITW-AITR-2026-001
- Version
- 1.0 · October 2026
- Pages
- 491
- Size
- 30.1 MB
- SHA-256
- 0e96be2e4685bd442178a9bb0860b345f3cd9109be6bce7d36d11bf68c082f50
AITW-AITR-2026-001
The TOPS Trap: How Edge AI Optimized for Token Efficiency—and Forgot to Engineer for Truth
A reproducible evaluation of source grounding, hidden system instructions, political-topic asymmetry and trust architecture in local AI.
What is inside
The paper gives the research questions and hypotheses stated before formal testing, the test environment and method, the grounding, hidden-instruction, topic and confidence results, the vendor correspondence record, a proposed Edge-AI Trust Benchmark, recommendations, and four appendices: the evidence and source register, the bibliography, the claim traceability matrix and the revision history.
Scope. Findings apply to the exact configurations tested, on one hardware platform and two topic packets. They are not claims about all local models or all edge AI. The Ministral claim scores are a manual judgement; every scored claim is in the evidence archive so others can re-score it.
Disclosure. AIToolworx has a commercial interest in AI architecture, security consulting and application of the trust-assessment methodology proposed in the report. The proposed benchmark is an AIToolworx framework and has not been adopted as an independent industry standard.
Prefer the short version? Read the executive summary.