Styxis
Company
Knowledge

A virtual PCB inspection experiment must keep synthetic and observed defects apart

Freeze the board revision and the origin of every defect label before testing an inspection idea.

An experiment can confuse its evidence

Synthetic examples can help prepare an inspection workflow before a physical board reaches a line. They can also make an experiment appear convincing if generated labels are mixed with observed defects. Truthound can check a proposed experiment ledger for the declared sample-origin and label policy. It cannot judge an optical defect or prove an inspection system improves manufacturing yield.

The manufacturing collaboration

Siemens announced a collaboration with NVIDIA on October 22, 2026 involving digital twins and synthetic data for PCB optical-inspection preparation. The release includes project targets for engineering effort, false calls and yield. Those targets should remain attributed goals. They are not observed results from a reader's board design or an independently reproduced benchmark.

Freeze the sample population

Choose one board revision for the first experiment. Give each sample an identity, origin, label, reference-annotation version and split assignment. Real observations and generated examples can both be useful, but answer different questions. A generated defect label may test whether a pipeline preserves a known annotation; it does not establish how frequently that defect occurs in production.

Keep origins explicit

For the observed and synthetic sample inventory prepared as a Polars DataFrame, Truthound Core Python SDK 3.3.4's InSetValidator can compare sample-origin values with an owner-supplied list. An explicit pair rule can further distinguish which label categories are admissible for each origin. Define missing labels separately. The policy should expose an unknown origin rather than silently assigning it to real observations, since that would change the interpretation of the experiment.

A comparable Pandera policy

Pandera 0.35.0 offers schemas and custom checks for the same prepared-data policy, including Polars. Keep an existing tested origin check when its split and label transformations are understood. Neither library's presence validates the synthetic generator or the inspection model. The choice is how to maintain a visible acceptance contract around the experiment, not which tool can claim the manufacturer's target benefits.

Create one informative failure

Make the first failure informative: insert a synthetic sample whose origin is marked observed. The test should reject or flag that inconsistency under a defined rule. Then ask an inspection specialist whether the labels and board revision are a meaningful test population. These are different reviews; a structural check cannot replace the domain judgment.

Preserve the cohort when designs change

Compare outcomes only after the population and metric definitions are fixed, keeping generated and observed results distinguishable. If a new board revision arrives, begin a separately identified cohort rather than extending the old one unnoticed. This preserves a small experiment's learning value without converting promising preparation technology into a claim of measured production improvement.

Evidence

Event checked
Subject checked
Modified