Not every check
means the same thing.
Independent checking is the family principle. The property established by the check is specific to the tool and the experiment.
Explore all 15 toolsA satisfying assignment is not the same as a proven optimum
SATForge rechecks clauses for a returned satisfying assignment. bnb_certified can close the supported knapsack proof to gap zero. Sinkhorn marginal checks assess transport feasibility, not unregularized optimality.
A residual or tolerance tied to the original problem
krylov_scale recomputes a true residual. SpectralForge uses round-trip, energy and reference checks. TomoForge checks data fit. Residual size, forward error and reconstruction quality are different properties.
The evaluated dataset and model still matter
ANNForge measures recall against exact-search references. GraphForge uses algorithm-specific graph checks. enkf_scale reports comparison with an exact Kalman-filter reference on its qualification cases.
Coverage is probabilistic, not an absolute promise
certmc and certcount report uncertainty intervals under their method assumptions. sketchstream reports sketch-specific error behaviour. An interval contains the truth with its stated coverage, not with certainty on every run.
The GPU path.
And the path actually tested.
Forge15 targets fast GPU execution with a separate checker. The supplied campaign did not exercise TomoForge’s matrix-free path on GPU; do not read the family label as a claim that it did.
A different package.
The same precision.
Explore source-bound integrity, candidate and mathematical-completion checks in UNISOLVE. These do not replace Forge15’s tool-specific exact, numerical and statistical checks.
Read the verification guide