A Principled AI Path to Spec-Driven Verification

In semiconductor design, specifications are often seen as the ultimate reference point—the guiding document that defines what a product must achieve. Yet in practice, specifications are rarely final or unambiguous. They evolve constantly under the pressure of changing market demands, design constraints, and rapid engineering iterations. For verification teams, this presents a major challenge: how to consistently generate accurate, production-quality test plans from shifting and sometimes contradictory specs.

This challenge is now at the center of industry discussions around AI in verification. Can AI, particularly large language models (LLMs), read specs and automatically produce detailed test plans? Conceptually, the idea is appealing. In reality, accuracy and reliability remain significant hurdles.

The Problem with Specs

As Adnan Hamid (President & CTO, Breker Verification Systems) points out, real-world specs are not neat, finalized documents. Instead, they are layered with Engineering Change Notices (ECNs), incremental updates, and multiple product variants. Each ECN often assumes deep prior knowledge, making clarity a secondary priority.

This creates several challenges:

• Human error: Even experienced engineers can misinterpret evolving specs.

• LLM limitations: General-purpose AI models may lack the specialized domain expertise to interpret subtle technical details.

• Context gaps: Expert verification engineers supplement written specs with insights from architects, designers, and years of hands-on experience—knowledge that is rarely captured in documents or datasets.

A good example is the RISC-V specification, which evolves with a rich ecosystem of extensions and updates. These layers can sometimes create multiple plausible interpretations, leading to confusion even among experts.

Why AI Alone is Not Enough

The idea of LLMs parsing specs and directly generating verification test plans is compelling, but accuracy is not yet at the level required for production-quality signoff. In verification, “close enough” is unacceptable. A single missed or misinterpreted requirement can lead to design flaws, silicon re-spins, and costly delays.

This is why Breker Verification Systems has adopted a more measured, principled approach to spec-driven verification.

Breker’s Two-Pronged Strategy

1. NLP-Based Spec Interpretation

Breker is building an NLP-driven system—not LLM-based—that interprets specifications and converts them into Breker’s graph-based test synthesis platform. Unlike general-purpose LLMs, natural language processing (NLP) in a constrained technical domain can provide greater determinism and reliability.

From this foundation, Breker can generate comprehensive test plans covering:

• Debug and coverage support

• Emulation portability

• Security root of trust checks

• RISC-V compliance and SoC readiness

• Power management verification

This approach builds directly on Breker’s existing library of verification functions, developed and refined over years of production use.

2. Strategic Partnerships with Agentic AI Platforms

While NLP serves immediate needs, Breker is also partnering with agentic AI ventures that are experimenting with LLM-based spec interpretation. These ventures bring strengths in AI-driven automation, while Breker contributes deep expertise in verification and deterministic test synthesis. Together, they aim to bridge the gap between LLM-powered interpretation and production-quality verification outputs.

Why This Matters

The future of verification will not be about fully replacing engineers with AI. Instead, it will be about AI-augmented verification, where machines handle the heavy lifting of parsing and structuring complex specifications, while experts provide judgment and oversight.

Breker’s strategy reflects a broader principle across electronic design automation (EDA): successful tools must be grounded in a principled foundation. Just as EDA relies on principled simulation, principled place-and-route, and principled multiphysics analysis, spec-driven verification must also be principled to ensure accuracy and trust.

Conclusion

The excitement around AI in verification, especially LLMs, is well deserved, but accuracy remains the critical barrier. Breker’s two-pronged approach—NLP for near-term reliability and LLM partnerships for long-term potential—represents a pragmatic balance.

For the semiconductor industry, this is not just a technical shift but a cultural one: moving from purely manual spec interpretation to a hybrid of AI-driven automation and expert oversight. As these methods mature, verification teams may finally have the tools to keep pace with the relentless evolution of specifications.

Source - Semiwiki