Durable result evidence
Large certificates are stored as durable files; the response returns a digest and verification reports. They can be inspected independently of the selecting engine.
Integrity, checked candidates, optimality, infeasibility, unboundedness and original-source trust are distinct. The current release keeps those meanings separate.
Six illustrative explanations—not live solver output, a certificate upload or an exact schema for every profile.
integrity_verifiedThe retained payload matches its recorded identity.
solution_verifiedA candidate passes its applicable original-model checks.
optimality_verifiedA supported independent optimum claim passes.
infeasibility_verifiedThe supported original-model infeasibility claim passes.
unbounded_verifiedThe supported original-model unboundedness claim passes.
validThe performed checks find no contradictory claim; this is not synonymous with solved.
Original-source and trust flags remain separate as well. A valid-but-incomplete state is not admitted as a complete answer when --require-optimal applies. Exact field combinations depend on the selected result and profile.
Every accepted solve-any answer receives a solver-specific check and generic certificate verification.
Large certificates are stored as durable files; the response returns a digest and verification reports. They can be inspected independently of the selecting engine.
Supported irrational optima retain algebraic witness, objective and bound fields. Ordinary numeric fields stay null rather than misrepresenting an irrational value.
A specifically configured SCIP/VIPR route supplies original-model certificates for exact linear MIP. Ordinary SCIP, HiGHS or other numerical statuses are not promoted to independent proof.
A parent bound transfers only after independent replay and preservation of every parent domain, objective and row. A fresh current-original witness must attain it before a new optimum is granted.
No. It restricts acceptance to an independently checked complete decision or global optimum. Unsupported cases, exhausted budgets and unknown outcomes retain their boundaries.
No. Frontends admit defined subsets. Optional reader output and high-level compiler translation retain their own trust boundaries.
No. Fixed-outside regional optimality applies only to that restricted problem unless an original-model global proof independently closes the claim.
No. Optional integrations are explicit. GCG decomposition/column generation, Concorde, Gurobi and Kissat are not native UNISOLVE engines in this update.
No. The customer must validate the model and interpretation of its requirements. Proof acceptance covers the admitted original model and theorem.
Keep a feasible solution, a checked bound, a regional optimum, an epsilon gap and a complete original optimum separate. Infeasibility and unboundedness have their own proof obligations. A checksum or a numerical backend status grants no global proof authority.
Retained independently checked original-model result with an explicit 20-million-work checker allowance. A larger checking allowance is part of the disclosed result.
Fresh proof tests block solver constructors, while hostile certificate changes are tested separately. Accepted solve_any answers need both solver-specific and generic certificate verification.
--require-optimal accepts only a checked complete decision or global result. It cannot force an unfinished search to produce one.
Start with the model, the environment
and the evidence your workflow needs.