START HERE / RUNNABLE PREVIEW
Start with the simplest engineering rule. Expose a direct-label shortcut, hold out complete converter designs, compare the same errors, and state exactly what the second result cannot prove.
Synthetic teaching data only. Not target-user tested. Not hardware validation, design approval, or evidence of transfer to your converter.
EVIDENCE CONTRACT
They are used as evidence states, not decorative emphasis.
BEFORE YOU START
The route supports a programming novice without simplifying the power-electronics context.
Do not upload confidential designs, proprietary simulator models, credentials, export-controlled material, customer data, or personal information to a public notebook or AI service. Use an approved local environment when the data requires it.
CHOOSE YOUR SUPPORT LEVEL
Both routes use the same files, grouped split, metrics, and evidence boundary.
GUIDED
Follow what to install, what to run, what output to expect, and why each check exists.
COMPACT
Inspect the manifest and notebook contract, reproduce the run, then move directly to the held-out result and limits.
PUBLIC STARTER KIT / V0.2.1-PREVIEW
The package is public and requires no email. It contains the exact two datasets, executed notebooks, runner, checklist, manifest, seeds, predictions, expected outputs, provenance, and licence.
ZIP SHA-2563ec6b01c72118bdbec60cf5516dda55ceb5d66ecfa7f384deb8638211d4df21a
Exercise 1 exposes direct-label leakage. Exercise 2 uses noisy proxies and a hidden synthetic reference, but still cannot estimate hardware performance or approve a model.
LEARNING PATH / 00–07
Each module names the decision and the artifact you should carry forward.
Name the target, grouping unit, error consequence, and simplest rule that could solve the task.
OUTPUTDecision brief + rule baseline
State what one row and one design represent, where each quantity came from, and its unit.
OUTPUTSchema + design_id
Check whether a feature contains or exactly reconstructs the target before treating performance as learned transfer.
OUTPUTLeakage audit + USE RULE decision
Keep every operating point from one design in one partition and fail if any design_id crosses the boundary.
OUTPUTGrouped split + overlap check
Evaluate each method on the same held-out rows; complexity must improve the decision, not merely a score.
OUTPUTFrozen comparison
Inspect precision, recall, F1, support, balanced accuracy, and the confusion matrix.
OUTPUTFalse-negative / false-positive review
Freeze the pipeline and evaluate on design groups withheld from training.
OUTPUTHeld-out result + scope
Record missing physics, unfamiliar inputs, validation still required, and when the system returns UNKNOWN.
OUTPUTBoundary + further-validation plan
CHECK UNDERSTANDING
The core route is complete when you can explain the split, result, and boundary. Transfer to a second authorised dataset is a separate extension.
Because the target is exactly vout_min_v < 4.90 V. The explicit rule reproduces it perfectly; logistic_regression adds an error and no decision value.
Because related rows from one design can appear in both partitions, allowing the model to exploit design identity instead of transferring to an unseen design.
The majority NORMAL class can dominate accuracy while the critical COUT_TOO_LOW class is missed. Critical recall and the confusion matrix expose that failure.
Only that the package reproduces two frozen synthetic decisions: use the exact rule in exercise 1, and continue validating the ML candidate in exercise 2. It does not prove causal diagnosis, transfer, or safety.
COMPLETION CRITERIA