Bring the original
model.

Validated typed intermediate representation, including symbolic theory and quantifier fields.

The supplied typed model must pass validation.

Mixed linear objective, graph-cut rewards, hard/soft clauses, linear constraints and supported finite globals.

Composition retains the accepted original constraints and objective.

1-based weighted edges, signed weights, parallel edges and self loops.

A graph-domain override must agree with the parsed model.

CNF clauses, empty and unit clauses, strict counts and zero terminators.

SAT witnesses and UNSAT refutations have different proof obligations.

Legacy top-weight and modern h hard-clause syntax.

Hard and soft constraints are kept distinct.

Linear Boolean constraints and objectives, including complemented literals.

Boolean linear scope; not an arbitrary integer-program parser.

Finite MIP/LP and exact rational linear model paths; configured SCIP/VIPR original-model certificates for exact linear MIP.

Broader source constructs depend on installed readers. Ordinary numerical status remains distinct from independently checked exact proof.

Supported finite constraint models; matching, flow and automaton CP globals are available within the admitted native inventory.

An emitted-model proof does not establish equivalence of a high-level MiniZinc translation. Confirm the actual construct subset before import.

Supported finite CP and scheduling, including variable quantities and exact end equations; native scheduling admits typed optional presences and variable durations/demands.

Only the admitted finite constraints, globals and scheduling formulations; not complete XCSP3 or every CP global.

TSP/ATSP inputs, including a matching/branch exact route admitting up to 256 cities by default; sparse directed arcs and signed binary-rational costs are admitted.

The 256-city limit is an admission guard, not a solve-time guarantee. Missing arcs are forbidden. Held–Karp retains its 20-city guard.

Supported graph normalization, including weighted METIS resource data and binary multi-resource balance.

Current G adds labelled multiway partitioning with per-part quotas and admitted proofs; unrestricted completion and GPU performance remain unestablished. Confirm the selected graph format and resource interpretation.

Preserved quantifier prefix with admitted quantified learning and prefix-preserving composition.

Original-game strategy checking is separately budgeted; finite decomposition does not flatten quantifiers.

Feature-based native routing for admitted affine, Boolean/integer, UF/array, bit-vector, guarded nonlinear and quantified reasoning, including multivariate Presburger elimination.

Full language conformance and arbitrary theory combinations remain unfinished. Optional Z3 is a separate external integration and trust boundary.

Supported exact rational normalization, dependency-free rational LP and supported polynomial/mixed-integer proof paths, including bounded multivariate spatial covers.

General nonpolynomial MINLP and exact irrational multivariate elimination are unsupported. Optional backend status is not an exact proof.

Source-preserving optional SCIP reader adapter.

Original bytes are retained for installed reader/backend integrations. No general native proof of original-language optimality is claimed.

Total parity games using the largest recurring priority convention.

Strategy checks use the admitted parity-game convention.

Boolean covering MIP.

The supported set-cover input, not every OR-Library problem family.

Undirected nonnegative-cost Steiner tree represented as an integer flow model.

The stated undirected, nonnegative-cost scope applies.

Classical job-shop input. Separately, the native scheduling scope includes admitted calendars, typed optional task presences and variable durations/demands.

Do not infer that every scheduling construct is expressible through this text format. Check the selected frontend and calendar/language limits.

Compile a high-level model with its data and includes, retain compiler provenance, and strictly parse the emitted FlatZinc.

A real MiniZinc installation is optional. High-level translation is trusted: native optimality on the emitted model is not a proof of high-level compiler equivalence.

Supported set-cover solving.

GCG decomposition and column-generation integration are not implemented.