A solver that agrees with itself has proved nothing.

Self-convergence, grid refinement and CPU-to-GPU parity all test consistency. None of them can tell you the method is right, because a wrong method converges just as cleanly as a correct one.

Consistency

is what self-convergence measures. A method can converge beautifully to the wrong answer, and the convergence rate will look excellent while it does. Validation methodology · our position

Independent target

is what validation requires - a reference the code did not produce and was not fitted to. Analytic zeros and published benchmarks qualify; your own previous run does not. Validation methodology · our position

Fitted is not predicted

an empirical potential tuned to reproduce a lattice constant will reproduce it. That is the design point, not a validation. Validation methodology · our position

Arithmetic is not physics

the trust verdict bounds how much numerical precision survived a computation. It says nothing about whether the model describes nature. GeoNum trust index · our position

Every solver here is checked against an analytic solution or a published benchmark, and where a check is only a consistency check it is labelled as one. That labelling is the difference between a validation suite and a dashboard.

6 tools, and the failure each one prevents

Six that close the biggest gaps. 54 in the full computational science and physics library.

Euler 3D Taylor-Green (WebGPU) HPC

The gap it closes. Taylor-Green has an analytic early-time solution, so the error is measured against something the code did not produce.

See it run in your browser →

Weyl psi4 Mode Decomposition HPC

The gap it closes. Spin-weighted spherical harmonic decomposition checked against the known mode structure, not against a previous run of itself.

See it run in your browser →

MHD HLLD Riemann Solver HPC

The gap it closes. Standard shock-tube problems with published reference solutions - the classic discriminating test for an MHD solver.

See it run in your browser →

Poisson Multigrid HPC

The gap it closes. Manufactured solutions give an exact answer to compare against, which turns a convergence plot into an error measurement.

See it run in your browser →

Topology Optimization (SIMP) HPC

The gap it closes. MBB, cantilever and L-bracket benchmarks carry published compliance values that nothing in the code was fitted to.

See it run in your browser →

GeoNum Cancellation Demonstrator STD

The gap it closes. Shows where naive floating-point accumulation stops being trustworthy, which is the mechanism behind every trust verdict on this platform.

See it run in your browser →

These are research and teaching solvers that run in a browser as the stage before a cluster, not replacements for production HPC codes or for the accepted tool in a regulated submission. Each states what it was checked against, and where a check only establishes internal consistency it says so rather than presenting it as validation.

Run one on your own numbers

Every tool above runs entirely in this browser tab. Nothing is uploaded, nothing is sent to a model, and each result cites the rule it applied.

Free tier, no card, no install. An account saves your work and lets you export a citable record of a run. See all 54 computational science and physics tools.