SiliconShield

SiliconShield brings Meta Muse into the semiconductor industry to help engineers investigate failures before chip fabrication. I built SiliconShield from my experience in semiconductor design and verification. Using Muse Spark 1.3 and Muse Wonder as research and coding collaborators, I constructed datasets, ran GPU training experiments, investigated learning failures, and built the assessment and training infrastructure. The investigation workflow uses locally served Muse Glimmer with two specialist skills: Contract-1 for register protection, privilege checks, locks, and address qualification, and Contract-2 for focused FIFO, counter, and arbitration behaviour. Muse Glimmer examines the RTL under the supplied assumptions, while Muse Spark 1.3 interprets the output evidence into a finding with supporting logic, disagreements, and open questions. As an additional research contribution, I fine-tuned Muse Glimmer 30B to perform the three-way decision task directly: satisfies the requirement, violates the requirement, or insufficient context. One published LoRA adapter covers both domains, achieving 97.22% accuracy on local-guard development examples. The submission includes reusable Contract skills, an investigation infrastructure, a working demo, a published fine-tuned Muse Glimmer adapter and scientific documentation. Together, these provide a practical foundation for AI-assisted hardware investigation, with consequential findings taken forward to engineering review, simulation, or formal verification.