MODEL CONTRACTS AND NUMERICAL VALIDATION
All inputs in SI except explicitly marked Wh/h dispatch quantities. Source: src/maxwellx.py, version 1.0.0. Read limitations before porting.

A. Ideal parallel plates
C0=pi^2 hbar c/720; U=-C0 A/d^3; P=-3 C0/d^4; F=P A; dF/dd=12 C0 A/d^5.
Zero temperature, perfect conductors, vacuum, negligible edge effects. Negative pressure means attraction. Differentiating U with respect to d recovers F. No power is calculated from a static gap. Independent small Gaussian input uncertainty gives sigma_F/|F|=sqrt((4 sigma_d/d)^2+(sigma_A/A)^2). Covariance and model error are excluded; inputs above 10% relative uncertainty are rejected. Mechanical stability output is only a local spring-gradient condition and assumes an equilibrium exists.
At 100 nm pressure is approximately -13.0013 Pa. At twice the gap its magnitude is divided by 16. Tests compare derivatives and analytic scaling, not only duplicated code expressions.

B. Ideal sphere-plane PFA
F=2 pi R (U/A)=-pi^3 hbar c R/(360 d^3). Restricted d/R <= 0.01. Real material corrections and quantified truncation error are absent. Arbitrary 3D arrays are not solved. External results require solver version, geometry, material provenance, convergence description, units and ordered numeric points. Import success never changes validation status to 'verified'.

C. Long-patch electrostatic limit
Pressure magnitude=epsilon0 Vrms_difference^2/(2 d^2). Vrms is the RMS difference between facing surface potentials in a long-correlation-length limit, not automatically the RMS on one measured surface. This is not a spectral patch model and should not be used to automatically subtract backgrounds. The 10 mV curve in the figure is an assumption.

D. Finite-temperature planar Lifshitz pressure
Identical local isotropic nonmagnetic semi-infinite half-spaces in vacuum. Let xi_n=2 pi n kB T/hbar, zeta_n=2 d xi_n/c, y=2 q d, w=sqrt(y^2+(epsilon(i xi_n)-1) zeta_n^2).
r_TE=(y-w)/(y+w); r_TM=(epsilon*y-w)/(epsilon*y+w).
P= -kB T/(8 pi d^3) sum_n' integral_{zeta_n}^infinity y^2 sum_p [r_p^2 exp(-y)/(1-r_p^2 exp(-y))] dy.
Prime halves n=0. Ideal r_TE^2=r_TM^2=1. Drude epsilon=1+omega_p^2/[xi(xi+gamma)], with n=0 TE=0, TM=1. Plasma uses gamma=0 and its finite TE zero-mode expression; it is not implemented as Drude with an arbitrarily tiny gamma. Constant dielectric uses n=0 TE=0, TM=(epsilon-1)/(epsilon+1). That dielectric scenario is a mathematical approximation, not a realistic all-frequency response dataset.
Quadrature and Matsubara truncation share y cutoff 50 by default. Re-run 40/60 and stricter quadrature tolerance for convergence. Returned quadrature error excludes truncation, material error and physical-model error. The included report checks a Drude scenario at 100 nm, 300 K, omega_p=1e16 rad/s, gamma=1e14 rad/s. These are hypothetical parameters. No gold label is permitted.
Analytic gates: ideal low-T/small-gap limit; ideal high-T pressure=-kB T zeta(3)/(4 pi d^3); Drude high-T half of ideal; no dielectric contrast gives zero; increasing plasma frequency approaches ideal. Positive T only; >10000 required terms fails explicitly. Nonlocal response, films, substrates, roughness, magnetic materials, curvature and arbitrary patterns are excluded. Call this numerically checked analytic-response solver, not a measured-material validated solver.

E. Energy and momentum
Residual=inputs-outputs-losses-DeltaStored. Named terms must cover one boundary and interval; null produces incomplete accounting. A negative DeltaStored is discharge. Zero balance does not prove all terms were measured. Three-sigma flag is only meaningful when the supplied combined standard uncertainty is justified, including covariance. Pure collimated emitted-radiation recoil is P/c; incident reflection is a different system. At 1000 W emitted power, recoil is 3.33564 microN.

F. Photon explorer
Pump f=f1+f2; photon E=h f; mean thermal occupation=1/expm1(h f/kB T), zero at T=0. No assumed conversion efficiency or flux prediction. Thermal occupation does not separate measured DCE photons from leakage or amplifier noise.

G. Helios CSV kernel
Canonical headers: time_s,voltage_V,current_A,run_id,condition. Additional columns retained only outside this analysis kernel in the original file. Strictly increasing elapsed seconds; one run per file. Hash original canonical text; retain the separately hashed instrument original when mapping units. Voltage and current must share a passive load and sampling epoch. For biased half-cells, integrated V*I is a terminal diagnostic and never delivered_load_energy_J. The UI must not provide passive-load confirmation by default.
Trapezoid integration excludes intervals touching missing samples, crossing condition labels or exceeding max_gap_s. It does not bridge outages, interpolate between runs or smooth. Coverage is covered duration/total span, not a count of samples. Negative powers are preserved; wiring sign conventions need review. This kernel does not compute biological attribution, confidence intervals or metrological uncertainty. High coverage does not eliminate aliasing; choose sampling based on instrument response and signal bandwidth.

H. Tank-to-home dispatch
Incident energy=irradiance * illuminated projected area * interval. Conversion efficiency is assumed, never inferred from tank volume. Direct DC generation serves auxiliary demand and then AC household demand through inverter. Excess charges battery subject to energy/rate limits; deficits discharge it. Charging and discharging are not simultaneous in an interval. Curtailment is explicit. Round-trip efficiency is eta_charge * eta_discharge, not either factor alone. Rate limits are at DC bus. No standby inverter drain beyond the specified auxiliary demand. No battery aging, temperature dependence, self-discharge or biological feedback. Piecewise-constant profiles require finer resolution for transients. Residual closes the electrical balance; incident optical energy minus generation is outside that electrical balance and remains a conversion loss.
Eight-foot diameter is 2.4384 m; top aperture is about 4.67 m2. Total inner electrode surface is not additional external solar aperture. A peak paper power density cannot be multiplied by arbitrary internal area to make a household-energy claim.

I. Communications
rho is a normalized Hermitian positive 4x4 density matrix; Alice indices come first. Local Kraus operators obey sum K†K=I. Bob is obtained by tracing over Alice. The before/after marginals agree for complete local channels. A postselection is not trace preserving and needs separately transmitted outcome information. Included tests use random mixed states and amplitude damping, not only a Bell state.

J. Classical temporal LC model
q'=i; L i'=-Ri-q/C(t); C(t)=C0[1+m sin(2 pi fm t)]. Energy=q^2/(2C)+Li^2/2. Modulation power=-q^2 C'/(2 C^2), resistor heat rate=R i^2. Therefore DeltaEnergy=modulation_work-heat. The numerical model integrates all terms and verifies the residual. It has one classical mode, no vacuum noise, no spatial propagation, no input source and no photon statistics. It is a preparatory energy-accounting exercise, not a replication of R02 or R09. A static impedance benchmark is also provided; it must never be labeled temporal reflection.

DEPENDENCIES AND REPRODUCIBILITY
Pinned runtime versions in requirements.txt are the versions actually used here, not claims about the latest releases. Run python run_validation.py from the package directory; Node must be available for browser parity tests. Tests use Python unittest. Figures and fixtures are rebuilt deterministically. See results/validation-summary.json and test logs. Browser tests execute ES modules under Node; live-site integration, accessibility and visual browser tests remain for the MaxwellX coding AI.