looptimum Black-box optimization for expensive evaluations Discuss a pilot

Case studies

Results are only useful when their evidence boundary is clear.

Explore documented engineering and model-training campaigns. Each case separates the observed outcome, the validation performed, and the limits on what the result establishes.

Read the case studies Discuss a pilot

Two campaign records

Two applications, with distinct evidence boundaries.

The engineering example includes downstream physical validation. The anonymized ML example demonstrates a compact guided training-recipe campaign and is presented with its small-budget limitations.

Bar chart comparing solver wall clock for the fine reference and validated coarse mesh

Validated engineering campaign

Fine-to-coarse simulation mesh search

A bounded engineering campaign paired its selected candidate with downstream solver and mesh-independence validation.

72.9% Fewer cells91.0% Lower wall clock11.1x Speedup

Domain-specific acceptance required a separate downstream solver pass.

Read the case study
Normalized held-out loss across ten evaluations in an anonymized GPT training campaign

Anonymized ML campaign

Guided GPT training-recipe search

Ten bounded evaluations tested four generic recipe controls while the training evaluator remained externally owned.

10/10 Successful evaluations1.02% Lower held-out loss~25% Fewer parameters

Single-seed, small-budget campaign.

Read the case study

Continue

Have an expensive evaluator of your own?

Review how the controller works, then decide whether a bounded and auditable campaign fits your workflow.

Discuss a pilot See how the loop runs