E-Sharp's Probe Makes AI Cite Its Evidence or Stay Silent in PCB Review

A single digital isolator, powered at 5 V while its bus ran 3.3-V logic, decided the fate of one board. Every number needed to catch the error sat in the schematic and the datasheet before release. Nobody had compared them.

That failure, described by Daniel Rhodin, founder and CEO of Swedish engineering firm E-Sharp AB, is the opening argument of a detailed essay on AI-assisted design review. The mistakes that trigger late-stage respins are rarely exotic. They are ordinary relationships hidden across schematics, bills of materials, datasheets and documentation.

Why do ERC and DRC miss these defects?

Electrical rule check catches unconnected pins and conflicting drivers. Design rule check confirms a layout obeys geometric and manufacturing constraints. Neither verifies design intent. A schematic can be electrically consistent while using the wrong component variant. A power rail can connect correctly yet fall short on current. A signal can land on a valid input in the wrong voltage domain.

Rhodin argues the human limitation is attention, not competence. No engineer holds every component variant, pin function, footnote and electrical limit in memory at once. Attention drifts toward complex power circuitry and FPGA configurations, while a quiet voltage mismatch passes unnoticed.

Why can't a general-purpose language model do it?

The tempting approach — export the schematic, attach the BOM and datasheets, ask "do you see any problems?" — works in demos and fails on commercial designs. Rhodin names three limitations.

• Context. A large design can exceed a one-million-token context window several times over, across thousands of nets and tens of thousands of pins. More prompt data can make answers worse, because relevant evidence drowns in plausible noise.
• Datasheet structure. Meaning lives in the relationship between heading, table row, column, unit, test condition and footnote. Text extraction can destroy that structure, letting the model pair a real number with the wrong parameter.
• The compulsion to answer. Generative models are trained to produce output even when evidence is incomplete. In Rhodin's testing, answers occasionally described the wrong component, package or variant — a 5.5 V absolute maximum instead of 6 V on a related part. The explanation still sounded coherent.

That last failure mode is the dangerous one, he writes. The source is real, the number is real, the reasoning sounds reasonable. The relationship between them is simply wrong.

What does Probe do differently?

E-Sharp's answer is Probe, a review system that builds a structured model of components, nets, pins, variants and their relationships, then links every conclusion to its source. For each check, the system retrieves only the relevant design slice and document passages, giving the language model a bounded problem with traceable evidence.

Its core principle is blunt: cite or abstain. A citation is not text the model writes; it is a reference the server dereferences before accepting the statement. A connectivity claim must name a net that exists on the board. A datasheet claim must name a part whose document was actually acquired. Unresolvable references are rejected, with the remedy stated: fetch the datasheet before citing it.

The evidence hierarchy is strict. Model reasoning sits at the bottom and cannot promote anything alone. Ten engineers agreeing leaves a value at the unsourced-agreement tier; one manufacturer application note raises it two tiers. Consensus, Rhodin notes, is a weaker basis than a citation while looking stronger.

What did a real review produce?

On one of E-Sharp's own boards, Probe generated 16 rule hits. Analysis reduced these to six relevant observations, four genuine defects, two of them blocking. One rule fired four times: three false positives and one real defect. A rule hit, Rhodin concludes, is a starting point for analysis, not a finding.

Coverage is reported per check as covered, partial or blind — and, more uncomfortably, per error class such as voltage-domain mismatch, absolute-maximum violation, power budget, reset sequencing and thermal. Probe reports thermal blind on every board, permanently, so nobody mistakes silence for coverage.

Rhodin's verdict is restrained. AI stops being a guessing machine when the engineer can see what each answer rests on. The most valuable output is sometimes a citation, a documented gap, or a plain statement that the question cannot yet be answered.