Start with an engineering question you can audit: what does the supplied reset failure show, which specification rule applies, and what remains unknown? This section designs a guided regression assistant for RTL and DV engineers, with equivalent evidence habits for STA and physical design. It searches approved log snapshots, reads an approved specification excerpt, and retrieves approved run history. It cannot launch a regression, edit a file, or change a waiver.
Lesson | Concrete deliverable |
|---|---|
Turn a regression question into a small, reviewable evidence-gathering task. | |
Separate the decision maker, controller, tools, and evidence record. | |
Validate exact inputs and keep log contents outside the authority boundary. | |
Choose one bounded action at a time and stop with an honest status. | |
Preserve evidence, unknowns, and useful next checks without hidden reasoning. | |
Make denial, missing data, repetition, and deadlines visible outcomes. | |
Study a complete standard-library baseline and a separate model adapter contract. | |
Grade evidence quality, boundary enforcement, and useful abstention on held-out cases. |
The deliverable is an evidence packet with bounded tool calls, source references, explicit unknowns, and a stopping reason. You will learn which decisions belong in deterministic code and which could later be proposed by a model. No provider account, dependency, EDA license, real design data, or network access is needed to understand the baseline. Follow the lessons in order; the starter lab gathers the entire controller in one listing.
No. At most, executing the baseline would demonstrate controller mechanics on synthetic fixtures. Diagnosis quality requires separate evaluation of a real model and representative approved cases.