RKO documentation

RKO minimizes an objective over random-key vectors. You provide the instance data, a decoder and an objective; the library provides the searches and cooperating portfolio. Domain feasibility and final validation remain your responsibility.

This is the development site for main, not a versioned manual for every published package. Start with installation and version boundaries. The current limitations distinguish implemented behavior, alpha-release boundaries and scientific claims that have not been established.

Start here

1. Choose and install a version: source versus a published package, optional extras, platform artifacts and free-threading limits.
2. Run a first solve: one algorithm, explicit execution and result decoding.
3. Load your own instance: bind application data to ordinary callbacks and optionally check the final solution. No adapter base class is required.
4. Configure the solver: engine selection, execution, budgets, results and errors.

Choose a workflow

Goal	Maintained guide
Select an individual algorithm	Standalone solvers
Set fixed scalars or adaptive grids	Parameter reference
Run compiled, block or built-in objectives	Rust and C++ engines
Run from the terminal	Repository example runner
Understand workers, startup and resource limits	Process workers, initialization, native threads
Convert a bounded mathematical model	Model contracts, historical qualification
Check a final solution	Optional validation
Export offline analysis	Reports
Illustrate declared decoder structure	Static diagrams
Retain authored decisions from original evaluations	Observed decisions
Bind trusted callbacks to a private HTTP service	Service and scaffold
Explore bounded decoder generation and held-out studies	Experimental evolution, locked studies
Find the papers behind an algorithm, a component or an experiment	Scientific references

Understand the limits

Execution accounting, convergence guarantees, C++ compatibility and performance criteria answer different questions. Passing installation tests does not establish solution quality, universal thread safety, timing guarantees or general C++ parity. Identified measurements and their original provenance stay in the benchmark reference.

Extend and maintain

• Architecture: public contracts, source layout and enforced import boundaries.
• Extension guide: replace operators or implement an algorithm.
• Contributing: tests, quality checks, evidence rules and review expectations.
• Documentation site: local preview, publication boundaries and recovery.
• Controlled releases: owner-approved publication, not an automatic consequence of merging code.