Fastest and most certifiable are different questions.

Performance

Selects the route with the lowest machine-adjusted empirical predicted cost. It is the mode that wins the benchmark, with median regret 1.000 — on half the cases it picks exactly what an oracle with hindsight would have picked.

Its worst case is 4.210× the oracle. That number is the honest reason the other mode exists.

Assurance

Selects the route with the lowest route-conditional conformal upper cost bound. It costs 188.089 ms across the benchmark against performance mode's 99.995 ms, and its max regret is 6.745×.

Paying nearly twice as much to be slower sounds like a bad trade until the question changes. Performance mode answers which route is likely to be quickest. Assurance answers which route lets me state a cost I can defend. Only the second one can be put in a contract.

That distinction is the third of the four reasons this release is a candidate rather than a stable 1.0.

Batch

For small registries, individual GCD. For a candidate registry and batch that are both large enough, a shared remainder tree. The crossover is measured, not assumed — the project records probe cost at bounds from 97 up to 100,000.

At batch size 128 and bound 10,000 the calibrated engine and the measured best engine were the same. That agreement is recorded as a result, not assumed as a property.

What a mode does not decide

• Choosing a mode≠choosing an answer — both modes factor the same integer correctly
• A lower predicted cost≠a guarantee of a lower actual cost
• A conformal upper bound≠a simultaneous guarantee over adaptively compared routes
• Calibration agreeing with measurement once≠calibration being correct in general