solve-model

Solve Model

Save complete object-by-period and scenario results. Recompute feasibility from delivered decision tables; carry opening inventory, conversion and losses through every period. A timeout or incumbent is not a certified optimum. Compare plans on the same inputs, capacities, opening stock and risk definition.

Read references/figure-result-handoff.md when results feed specialized figures. Save complete object/scenario results, baselines, termination status, units, fitted curves and actually computed interval bounds. Renderers display these results; they must not refit models or manufacture uncertainty.

Turn an agreed route into results that can be explained, checked and used in a competition paper.

Workflow

1. Confirm each question's objective, variables, units, constraints, assumptions, output and upstream model-selection rationale.
2. Read references/reasoning-and-validation.md and the relevant part of references/algorithm-api.md. For inverse/physical models, uncertainty-aware decisions or dependent-data experiments, read the applicable section of references/research-routes.md; validate the route on real inputs and preserve its failure boundary.
3. Formulate the real relationship before choosing a solver. Use analytical derivation for conservation, geometry, feasible ranges, reduction, scale or monotonicity whenever available.
4. Implement a simple baseline before the complex route. Make data slicing, parameters, seed and units explicit; preserve failed runs and limitations. If a decision uses a model-generated forecast, create predictions with rolling origins using only observations available at each issue time, preserve the issued predictions, and estimate forecast-error or risk terms only from previously completed targets.
5. For a critical event, define the event function, locate a first bracket, refine it and save left/right states. For a search, establish its prerequisites before bisection or Brent.
6. Audit hard constraints independently. Compare with the declared baseline and, where risk warrants it, with an algorithm that does not share the same key implementation path.
7. Separate parameter sensitivity from numerical convergence. Explain whether a perturbation changes a decision threshold rather than merely reporting a changed digit. Predeclare final holdout or extrapolation checks and do not repeatedly tune against them.
8. Save structured results, source tables and validation evidence so figures and prose can be regenerated without copying numbers by hand. Preserve run identity, input/configuration versions, actual seeds, termination state and failed cases. Separate observed results, mathematical deductions and proposed explanations in the writing handoff; never report planned validation as completed.