A consistent entry point
CLI solve-any and MCP solve_any use the same problem-centered entry. The Python API and stdio MCP server operate without a display, browser, TTY or prompts.
One workflow.
Explicit limits.
Unified dispatch, broader admitted models, revision-aware reoptimization and scoped recovery. Use the exact runtime boundary your task requires.
Current source G: UNISOLVE 0.16.1, 82 profiles, 65 entry functions, 133 Python exports, 38 CLI operations and nine stdio MCP tools. All profiles were exercised through the installed-package campaign; optional declines and incomplete outcomes remain explicit. Current evidence ↗
ONE REQUESTChoose the task.
Choose the task.
Let the interface dispatch.
solve-any detects the file and dispatches or races compatible registered profiles. Profiles may share an implementation.
Checked acceptance
Compatible engines can run in owned processes. A candidate must pass its solver-specific and generic certificate checks before selection.
Seventy-eight profiles
SAT, weighted MaxSAT, PB, MIP/LP, CP/scheduling, SMT, QBF, TSP/ATSP, MaxCut, partitioning, parity games and polynomial MINLP are represented within admitted scopes.
CHANGING MODELSRevise the question.
Revise the question.
Preserve the evidence.
LinearSession supports added constraints, checked warm starts, source changes, snapshots and exact reoptimization.
Current-original export
solve_result() returns the current original Problem and standard SolveResult for ordinary certificate export.
Conditional bound transfer
Independently replayed parent bounds transfer only when every parent domain, objective and row is preserved. A fresh attaining witness is still required.
Safe update boundaries
Relaxing rows, domains or objectives invalidates transfer. Interrupted updates preserve the previous session atomically; optional cache-copy stops remain explicit.
Existing finite-domain/SAT and SMT session contracts are separate. The new LinearSession capability is not a promise that every source-backed mutation or live frontier is reusable.
SCHEDULING & EXACT MODELSRequirements can be
Requirements can be
decisions too.
Use the broader admitted model paths without losing the original constraints or exact interpretation.
Scheduling quantities
Native scheduling admits typed optional presences, variable durations and demands. The supported XCSP3 subset includes variable quantities and exact end equations.
Resources and calendars
Matching/flow/automaton CP globals and resource calendars expand the supported modeling options. The retained calendar and language limitations still apply.
Exact mathematical outputs
Supported irrational optima, one-Real nonlinear/quantified fragments, unbounded one-integer arithmetic and guarded polynomial combinations retain independent original-model checks.
CONTINUATION & RECOVERYRetain useful work.
Retain useful work.
Name the contract.
A warm restart, exact frontier continuation and durable cover recovery are different operations.
Generic and packed warm restart
The earlier documented generic and packed restart paths restore checked saved work; they are not blanket live-frontier serialization. The current update adds packed-bit neighborhood repair without expanding the whole assignment.
Exact continuation and CUDA covers
The separate exact Boolean frontier runtime retains its admitted scope. The current build adds supported exact CUDA regional search and guarded finite covers with durable distributed recovery.
Measured hardware recovery
Recorded local-CPU/existing-Brev-CUDA recovery passes at 10K, 30K and 100K variables with 30K, 90K and 300K edges. Recovered assignments exactly match uninterrupted assignments; checkpoint resume preserves the checked objective.
One physical CUDA GPU
The measurements use two physical hosts and one physical CUDA GPU. Edges are sharded; assignment/gain vectors are replicated. Physical two-GPU performance is not validated.
No universal learned-policy speedup
The router admits compatible capabilities and exchanges only checked candidates/theorems. A fitted selector without held-out gains is not automatically promoted; negative timing/work findings remain in the record.
Explicit dependencies and budgets
Optional extras include highs, scip, smt, cuda, packed, learning, mpi and test. A MiniZinc compiler is optional. Installing or running UNISOLVE does not provision hardware. Work/node/byte guards have their stated scopes.
MODEL COMBINATIONSCarry the requirements.
Carry the requirements.
Check the whole claim.
Weighted resource balance
Supported binary partitions account for multiple weighted resources. Graph-to-MIP transport and checked symmetry/count reductions preserve the original graph requirements. General multiway partitioning is not included.
Supported mixed proof paths
Heterogeneous polynomial components, integer branches and admitted bounded multivariate spatial covers can combine under original-problem checks. Arbitrary theories and nonpolynomial models are not automatically admitted.
Checked reuse, not inherited trust
Supported source translations and strengthened linear models can reuse independently checked bounds. SAT can incorporate admitted CP/MIP reasoning only when it contributes to a checked master proof.
LEARNING & RECOVERYUse measured choices.
Use measured choices.
Keep the limits.
Do not promote a failed experiment
Defaults were promoted only for MIP/hybrid on the measured Brev host. The newer partition selector missed its target; other hosts/domains retain their documented defaults.
Different engines retain different state
Incumbent snapshots, admitted exact frontiers, SAT cubes and spatial pending cells have distinct recovery contracts. Not every live LP or theory trail is saved.
Transport is not scaling proof
Local/SSH workers, admitted MPI portfolios, shards and recovery are implemented. Actual SSH MCP transport was tested. General cluster performance and physical multi-GPU scaling remain unverified.