Algorithm Performance ​

compat-finder currently ships with two built-in algorithms: binary-split and leave-one-out.

Here, "performance" mainly means how many test rounds the troubleshooting flow needs.

TL;DR ​

• In almost all cases, leave-one-out uses exactly n questions; it becomes n + 1 only when the final result covers all targets.
• binary-split varies with the size and distribution of the final result set and usually falls between about max(1, ceil(log2(n))) and 2n - 1 questions.

The crossover point is around 13.77%; below that, in most cases, the worst case for binary-split is also better than leave-one-out.

Conclusion: binary-split remains the default because, in real troubleshooting, problematic targets are usually a minority.

How to Choose ​

• You expect to end up with only a few incompatible targets: prefer binary-split
• You expect to end up with many incompatible targets: consider leave-one-out
• You are unsure: start with the default binary-split
• You want to see how to switch algorithms in practice: continue with CLI and API Reference

Overall Trend ​

The chart below summarizes the overall behavior across all non-empty result sets.

When the Final Result Has Only 1 Target ​

Single-target cases are one of the clearest situations where binary-split wins.

When the Final Result Set Is Small ​

At 12.5% of the total target set, binary-split has a clear advantage.

Near the Worst-Case Crossover ​

Around 13.77%, the worst-case behavior of the two algorithms starts to get close.

Near the Average-Case Crossover ​

Around 22.83%, the average-case advantage starts to flip.

When the Incompatible Set Is Large ​

When the final result is close to "almost everything is incompatible," leave-one-out becomes more consistent and often uses fewer rounds.