Can an AI-generated PCB survive engineering review?

The PCB Hardening Problem

Asking general-purpose AI agents to write a schematic or route copper is easy, but they design blindly. They produce plausible-looking circuits that ignore thermal limits, hallucinate connections, or miss critical safety interlocks.

To make AI-designed hardware safe and repeatable, Lyon Industries is building a deterministic validation pipeline. The goal is to compile plain-language intent directly into verified, manufacturable PCBs. If we can enforce engineering rigor automatically, hardware development becomes as fast, cheap, and agile as software.

The compiler does not replace professional design review. Instead, it converts manual checks into automated gates, verifying every component, part number, and operating envelope against primary source documents before any fabrication files are exported.

The Validation Harness

Our pipeline wraps AI agents in a strict engineering environment. It defines requirements, verifies component provenance, and runs layout checks inside a pinned KiCad 10.0.3 runtime.

Before drawing a single line of copper, the compiler verifies that every claim is backed by a registered manufacturer PDF. If a requirement is open or a calculation fails, the pipeline halts. The system acts as a strict check on agent autonomy, enforcing DRC, ERC, and release compliance at every compile step.

1. 01

   Requirements

   Nine project requirements and two decisions are hashed into a candidate baseline. Seven requirements are still open, including the load, field supply, channel count, environment, and fault behaviour.

   9 requirements / 7 open

2. 02

   Source and evidence records

   Source, claim, component, circuit-block, artifact, and review records retain part identity, usage rights, document locations, hashes, operating conditions, and review state.

   17 schemas / 105 tests

3. 03

   Agent skill

   One portable skill gives both agents the same requirements, source-capture, PDF-indexing, validation, release, and isolated KiCad commands.

   Codex / Claude Code

4. 04

   Native EDA

   KiCad 10.0.3 source and library revisions are pinned. A native passive fixture exists, but the source-built runtime has not completed qualification.

   Runtime not qualified

5. 05

   Manufacturing release

   The release model checks artifact roles, hashes, tampering, requirement baselines, and approval state. No irrigation fabrication package exists.

   No product files

The Bürkert Footnote Surprise

Our first roadblock came from a silent failure in AI reasoning, demonstrating why automated validation is necessary.

For a low-pressure gravity system, the agent selected the Bürkert Type 6213 (article 221630) solenoid valve. The shop page listed a pressure range starting at 0 bar and a flow coefficient of 8.3 m³/h, which looked perfect. But deep on page 18 of the datasheet, note 2 revealed that the valve requires a 0.5 bar differential pressure to open fully. Under gravity pressure (0.1 bar), it would choke the flow. The AI saw the shop values but missed the qualifier, a mistake that would have shipped non-functional hardware to a physical installation.

This is a component rejection, not a measured valve test. Product availability, field pressure, pipe loss, filtration, and delivered flow remain separate questions.

The Irrigation Test Vehicle

We are testing this compiler on a low-cost, offline irrigation controller. The goal is to prove the pipeline can design a simple, repairable board that survives real-world field conditions.

The reference requirements are based on a shared-tank distribution archetype in Tamale, Ghana, where women farmers manually water dry-season plots. An offline controller could automate water scheduling and allocation. However, field studies in Zimbabwe show that 84% of relief drip kits fail within three years due to a lack of local support. The PCB is justified only if it reduces lifecycle and repair costs compared to a passive siphon or manual valve.

Before releasing any board files, we are building a physical test bench in Stavanger to characterize our provisional SMC VXEZ23 solenoid load. We must measure the startup current, holding power, and inductive turn-off energy. The compiler's software checks are passing, but the ultimate proof must be physical bring-up.

Checks completed on 12 July 2026

The table records the latest local verification on 12 July 2026. Counts are useful only with their boundary: these tests exercise the engineering records and automation, not an irrigation circuit or manufactured board.

Adversarial review found that supporting artifacts and reviewer decisions need stronger hash binding. Those checks are still being added. Version 0.2 evidence records cannot yet approve a component or fabrication release.

Work remaining before schematic capture

The next work is measurement and component review. The schematic follows only after the load and supply limits are known.

1. 01

   Freeze the first physical scope

   Use one 12 V valve channel and a controlled low-pressure water fixture as the bench proof, or wait for a measured field installation. The bench-first path is the current recommendation.

   Decision

2. 02

   Bind the decisive source evidence

   Capture exact manufacturer bytes where rights allow, bind PDF and web locators, and attach reviews to immutable record and artifact hashes.

   Source review

3. 03

   Qualify the EDA runtime

   Complete the pinned KiCad build, then run the native fixture through ERC, DRC, schematic parity, Gerber, drill, BOM, placement, and drawing exports.

   KiCad

4. 04

   Measure the representative load

   Record startup current, holding current, turn-off energy, pressure, flow, leakage, temperature, brownout response, and closure time before choosing the PCB switch and clamp.

   Measurement

5. 05

   Design and release the board

   Create the native schematic and PCB, prove the project-specific checks with deliberate faults, and reconcile every manufacturing artifact to one reviewed source revision.

   Schematic + PCB

6. 06

   Manufacture only after approval

   A quote and upload rehearsal may be automated. Sending design data, paying a supplier, and ordering boards remain explicit owner decisions. Field claims wait for physical evidence.

   Order approval