Column: Silicon systems design
Figure 2: Not all AI verification is equal
This distinction also changes human
review. Engineers may spend less time manually launching tools, moving information between environments and searching through large log sets. More engineering effort may shift toward defining objectives, controlling assumptions, reviewing evidence, and deciding whether the residual risk is acceptable. The engineer therefore becomes less
of a workflow operator and more of an authority over intent, evidence and acceptance.
AlpinumDV provides one example of
this approach, combining AI-assisted verification activity with structured verification artefacts, controlled execution and reviewable evidence, whilst maintaining traceability between engineering intent, generated assets and tool results.
Avoiding circular verification One of the most important risks is circular verification. Suppose the same AI system interprets a requirement, generates an assertion, decides which test should exercise it and then judges whether the resulting evidence is sufficient. The workflow may look complete while all four steps share the same misunderstanding. Using a second agent doesn’t
automatically remove this problem if both agents rely on the same model, retrieved context or hidden assumptions. The strongest workflows deliberately introduce independent evidence. That may include deterministic or independently executed tools, golden reference models, separately authored properties, specification- linked acceptance criteria, differential checking, human review or multiple verification methods addressing the same risk from different directions.
14 October 2026
www.electronicsworld.co.uk The objective is not to prevent an
agent from iterating. Iteration is one of its strengths. The objective is to ensure each iteration is constrained by evidence; the agent can’t simply redefine its answer to support its previous answer. Assurance strengthens as a
workflow moves from model self- critique to executable external checks, independent engineering verification, and requirements-linked evidence interpreted through engineering judgement.
Measure evidence, not agent activity Agentic workflows can produce impressive volumes of activity: generated tests, assertions, tool calls, summaries and proposed fixes. Tese are poor success metrics on their own. Verification teams should instead
measure false acceptance, requirements- to-evidence traceability, reproducibility, review effort, confirmed defect discovery, time to validated root cause and how oſten an agent correctly escalates uncertainty rather than forcing a conclusion. A useful agent is not one that always
produces an answer. It is one that can recognise when the available evidence is insufficient.
From self-verifying AI to defensible engineering The direction of travel is clear. EDA agents are becoming capable of longer and increasingly autonomous engineering workflows. Continuous validation against engineering tools is an important step towards making that autonomy useful. However, “self-verifying” should not
mean that an AI system is allowed to define its own truth. The strongest architecture is one in which agents can plan, act, investigate and retry, while independent tools and controlled engineering artefacts continuously test their assumptions. Every important conclusion should retain its provenance: what requirement triggered the action, what data and revision were used, which tool produced the result, what assumptions applied and who accepted the evidence. AI agents may eventually execute
substantial parts of a verification workflow without continuous human intervention. Trust will not come from autonomy itself. It will come from an evidence chain strong enough that an engineer can understand, reproduce and defend the result.
This column continues in next month’s edition of Electronics World
Page 1 |
Page 2 |
Page 3 |
Page 4 |
Page 5 |
Page 6 |
Page 7 |
Page 8 |
Page 9 |
Page 10 |
Page 11 |
Page 12 |
Page 13 |
Page 14 |
Page 15 |
Page 16 |
Page 17 |
Page 18 |
Page 19 |
Page 20 |
Page 21 |
Page 22 |
Page 23 |
Page 24 |
Page 25 |
Page 26 |
Page 27 |
Page 28 |
Page 29 |
Page 30 |
Page 31 |
Page 32 |
Page 33 |
Page 34 |
Page 35 |
Page 36 |
Page 37 |
Page 38 |
Page 39 |
Page 40 |
Page 41 |
Page 42 |
Page 43 |
Page 44
