Ports That Compile, Run, and Quietly Lie

A port that compiles and runs looks finished. That's the trap. Getting a machine learning pipeline to build in a new language and produce output is the easy 80%. The last 20%, making sure that output actually matches the original, numerically, at every stage, is where most ports quietly go wrong.

This post is about the gap between "it runs" and "it's correct", and the discipline needed to close it when porting an ML pipeline from Python to C++.

Why ports fail silently

A ported pipeline that compiles cleanly and produces plausible-looking output gives false confidence. Floating point differences, subtly different default parameters, off-by-one errors in preprocessing, or a different random seed can all produce results that look reasonable but diverge from the original in ways that matter for downstream accuracy.

The danger is that these bugs don't crash anything. They just make the model slightly, silently wrong, and that kind of wrongness can sit in production for months before anyone notices a pattern.

Stage-by-stage numerical verification

The only reliable defense is verifying intermediate outputs at every stage of the pipeline, not just the final prediction. Preprocessing, feature extraction, normalization, model inference, and postprocessing each need their own checkpoint where the C++ output gets compared against the Python reference on the same input, with a tight numerical tolerance.

When a stage doesn't match, the bug is usually one of a small set of culprits: a default argument that differs between library implementations, a different order of operations that changes floating point rounding, an off-by-one in an indexing loop, or a subtly different interpolation or resampling method. Isolating the stage narrows the search dramatically compared to debugging from the final output backward.

Key takeaways

• Treat a port as unproven until verified stage by stage against the original implementation.
• Build your test harness before you start moving code, not after.
• And remember that a clean compile and a plausible-looking output are necessary but nowhere near sufficient for calling a port done.