Nabla is not compared against itself. Every case states its reference source - a closed form, a published benchmark, another code on the same model, or an internal invariance the other tiers can't see - and is judged against it with a tolerance fixed before the run.

Cases

V01 — Coaxial cable, long solenoid and current loop

Tier A - Analytic · worst judged error +1.01 % · 4 models, 52212 nodes, P1/P2 · 23.6 s

A round conductor, a thick solenoid and a current loop, judged against closed-form magnetostatics in both the planar and axisymmetric formulations.

V02 — C-core electromagnet with an airgap

Tier A+X - Analytic + Cross-code · worst judged error +2.62 % · 12 models, 62178 nodes, P1 · 116.2 s

A wound C-core across an airgap, nonlinear and saturated, judged against FEMM on the identical model and against Ampere's law on a contour in the window air.

V08 — Steady and transient heat conduction

Tier A - Analytic · worst judged error +0.86 % · 12 models, 41257 nodes, P1/P2 · 48.0 s

Six Fourier-conduction problems with closed-form solutions, each solved at both element orders on the uncoupled thermal module.

V09 — P1 vs P2 convergence study

Tier S - Self-consistent · worst judged error -0.082 abs · 36 models, 295843 nodes, P1/P2 · 104.3 s

Three exact problems, meshed at five densities spanning 16x in element count, checking that the solver's error falls at the theoretical order as the mesh refines.

V10 — Nabla against FEMM on identical models

Tier X - Cross-code · worst judged error +1.94 % · 6 models, 19408 nodes, P1 · 315.3 s

Four physics and six models built, solved and read in both Nabla and FEMM 4.2 inside one run - the comparison a prospective user makes anyway, done here with the models published.

V11 — TEAM Problem 30 - induction motor with an analytic solution

Tier A + B - Analytic + Benchmark · worst judged error +1.92 % · 9 models, 92606 nodes, P1 · 1127.7 s

The one rotating-machine benchmark with a closed-form answer: a three-phase exposed-winding induction motor, judged against TEAM Problem 30's own published solution.

V12 — Skin effect: round conductor, slab and slot conductor

Tier A - Analytic · worst judged error +1.32 % · 16 models, 50619 nodes, P1 · 52.1 s

Three one-dimensional eddy-current problems with closed-form solutions, spanning a wide frequency sweep on each geometry.

V13 — Magnetized cylinder and far-field truncation study

Tier A - Analytic · worst judged error +0.60 % · 10 models, 70192 nodes, P1 · 50.3 s

A uniformly magnetized cylinder solved with the outer boundary at increasing radii and with the far-field (balloon) boundary condition, against the closed-form exterior field.

V14 — Sector vs full model; sliding band vs conforming mesh

Tier S - Self-consistent · worst judged error +0.36 % · 16 models, 192337 nodes, P1 · 1229.0 s

One machine as sixteen models: periodic sector against full ring, and sliding band against a conforming mesh, checking that two constructions Nabla calls exact really are.

V15 — Force methods cross-check on a linear actuator

Tier A+S - Analytic + Self-consistent · worst judged error +1.40 % · 13 models, 117544 nodes, P1 · 72.7 s

A C-core pulling on a plunger across two airgaps, with force computed three ways - the eggshell region integral, a Maxwell-stress contour, and virtual work - and checked for mutual agreement.

V16 — Transformer: magnetizing and short-circuit impedance

Tier A + X - Analytic + Cross-code · worst judged error +2.43 % · 7 models, 50199 nodes, P1 · 1223.3 s

A two-winding transformer solved open-circuit and short-circuit, with its magnetizing and leakage impedance against closed-form winding theory and against FEMM.

V17 — Circular loop self and mutual inductance, axisymmetric

Tier A - Analytic · worst judged error -0.25 % · 3 models, 203724 nodes, P2 · 52.1 s

One circular turn asked for its own inductance and for its mutual coupling to eight coaxial neighbours, axisymmetric, against the elliptic-integral closed form.

V18 — Iron loss post-processing vs Steinmetz by hand

Tier A - Analytic · worst judged error +0.77 % · 11 models, 24520 nodes, P2 · 57.4 s

The iron-loss post-processor against the Steinmetz model its own coefficients are defined for. This case shipped FAIL, found three defects, and now passes.

V19 — Backward Euler vs Crank-Nicolson order study

Tier A+S - Analytic + Self-consistent · worst judged error -0.00 % · 24 models, 37140 nodes, P1 · 126.8 s

Backward Euler against Crank-Nicolson on transients whose exact answer is known, measuring the observed order of each scheme rather than asserting it.

V20 — Solver-option invariance matrix

Tier S - Self-consistent · worst judged error -0.79 % · 15 models, 86675 nodes, P1/P2 · 203.2 s

Half a dozen solver switches that are choices of method, not of answer - linear solver, time scheme, element order, band type, axial slices, field-circuit coupling - each moved to see whether the answer moves with it.

V22 — Linear machine and the linear sliding band

Tier X - Cross-code · worst judged error +0.40 % · 11 models, 473902 nodes, P1 · 400.2 s

A flat linear machine's force pair at eight mover positions, by three constructions and two codes, with the mover carried by the linear sliding band.

V23 — Loss to thermal handoff on a machine

Tier A+S - Analytic + Self-consistent · worst judged error +0.17 % · 3 models, 207608 nodes, P1/P2 · 52.5 s

The loss-to-temperature chain end to end: a machine's Joule and iron losses become heat sources in a nine-layer stack, judged against the layered-cylinder closed form and against conservation.