Reading an answer
Every number in an answer comes from a solver (a program that finds the best option) or the math kernel, never from a guess. Each answer carries a status label and a verification line.
Status labels
- Optimal. The solver proved that no better answer exists under your rules. For whole-number problems it also shows the gap (how far the best answer found could be from the best possible).
- Locally optimal. A good answer, but the problem has several "hills" and we cannot prove this is the highest. It is labelled "not certified as global" and, where possible, we show a bound (a limit nothing can beat) and the gap.
- Infeasible. Your rules contradict each other, so no answer satisfies them all. Example: x >= 5 and x <= 3. Check the rules for a typo.
- Unbounded. The goal can improve forever because a rule is missing. Example: maximize profit with no limit on production.
- Outside V1. The problem is too large or too specialized for the first version (complexity levels 8 to 10). ProvenDecisions says so and may offer a clearly labelled simpler version.
Verification
After solving, ProvenDecisions checks every rule against the answer (feasibility), checks that whole-number variables are whole, and, by default, runs an independent optimality check: the dual problem for a linear program, the optimality (KKT) conditions for a smooth problem. If a check fails, the answer is marked "Not verified" and is not presented as right.
What is proven: the label on every answer
Each answer ends with one short label that says exactly how much is proven:
- Exactly proven optimal. Linear programs. The walkthrough shows a step called Why this is optimal: a weight for each rule, so that adding the rules with those weights gives a limit no plan can beat, and the answer reaches it. Simple problems can be checked by hand. More complex problems can be checked with a spreadsheet. Every step is re-checked in exact fractions, not rounded decimals.
- Optimal, numerically checked. Convex problems where the optimality (KKT) conditions hold, or linear programs where the dual check agrees but an exact proof was not produced, for example very large models. Correct to many decimal places, but not an exact proof.
- Proven optimal by branch and bound. Whole-number problems, when you tick Show a checkable proof (integer problems) on the Solve screen. The walkthrough shows a step called Why this is optimal (branch and bound): a table of cases that together cover every whole-number plan. Each case is closed by weights on its rules, either showing it has no plan at all or giving a limit it cannot beat; every check is redone in exact fractions. Very large problems (over 400 cases or 10 seconds) keep the label below.
- Optimal within gap tolerance. Whole-number problems without the proof option. The solver's search proves the answer is best within a tiny tolerance, but that proof is not shown.
- Optimal per solver. The solver reports optimal and no independent check applied.
- Not proven optimal. A local optimum (a better answer may exist elsewhere), limited solver accuracy, or a nonconvex problem. We say so plainly.
Duals and sensitivity
A dual value (shadow price) tells you how much the best value changes if one limit grows by one unit, for example one more oven hour. A slack tells you how much of a limit is left unused.
Walkthrough depth 1 to 10
Depth 1 is the answer and a one-line formulation. Depth 3 (the default) is mostly plain-language: decisions, goal, rules. Higher depths add indices and sets, each equation, the code, solver settings, and how to read duals and slacks. Depth 10 is the full derivation. You can change it for any answer.
If an answer looks wrong
Most often a number in the problem was misread or a rule was left out. Look at the formulation ProvenDecisions shows, correct it, and run again. If it still looks wrong, tell us: support re-checks it with the solver and a person on our team reviews anything that is truly wrong.
Try ProvenDecisions free: type, photograph or say an optimization problem; every answer is checked by a solver.