Vibe Coding · Guardrails

Verify before you trust

AI-generated code fails differently from human code. It rarely crashes; it runs, looks professional and gives a wrong number. Engineering judgement is the guardrail, and a checklist makes it routine.

Checklist10 min read

How AI-built tools go wrong

Failure modeExampleHow to catch it
Plausible but wrong formulaBSF computed with the outer-race formulacompare with a published value for one known case
Silent unit errorsg treated as m/s², peak reported as RMSfeed a sine of known amplitude in known units
Off-by-one and resolution errorsFFT bin frequency shifted by one bincheck a tone exactly on a bin and one between bins
Hard-coded demo valuessampling rate fixed at 25,600 Hz in a hidden placerun a second file with different settings
Invented library functionsa function that does not exist in the library versionrun it; agents that can execute code catch most of these
Over-confident wording"Bearing failure detected" from one weak peakread every message the tool can show; ask for evidence and hedging
Data leakage in MLrandom window shuffle gives 99.8% accuracysplit by recording or machine; expect a lower, honest score

Verification checklist

Tick items as you go; your progress is saved in this browser.

Where AI-built tools fit, and where they do not

Good fit

  • Analysis and calculation aids reviewed by an engineer
  • Data preparation and reporting
  • Prototypes that prove an idea before a proper project
  • Training and teaching tools

Needs a formal software process

  • Protection, interlocks and safety instrumented functions
  • Control logic that changes equipment state
  • Regulatory or contractual calculations
  • Systems many people depend on without reviewing outputs

A tool can move from the left column to the right as it grows. When that happens, hand it to a software team with your specification and test cases. Those two documents are the most valuable part of what you built.

Key takeaways

  • AI-built code usually fails quietly, with wrong numbers, not crashes.
  • Known-answer tests, unit checks and edge cases catch most problems.
  • Protect confidential data and state clearly what decisions the tool supports.
  • Keep the specification and tests: they make the tool reviewable and transferable.