P=NP - Hyperbolic 3-SAT

Continuous relaxation on the Poincaré line. At ratio ≈ 4.267 (phase transition), IEEE 754 loses gradient signal to underflow - GeoNum preserves it.

GeoNum Precision Demo - MLE

Find the peak of L(μ) = Π_i P(x_i|μ,σ) via coordinate scan. IEEE 754 direct product underflows to 0 at large N - making every candidate look identical. GeoNum multiplies in log-space, staying exact at any N. No log-sum trick required.

Geometric Manifold Prediction Geodesic extrapolation · Jacobi field uncertainty

Module Ballistics Phase Diagram 2D FEM Geometric MD Helicity GHD Transport Batch Mode Plasma Medical Cosmos Advanced 3-Body

§ Formal Specification - v2.2

Approximation class. GeoNum is piecewise approximate with O(1/2048) quantization error per zone, drift-tracked. Each encode introduces quantization error ε_q ∈ [0, 1) shade units (systematic truncation). Under zone-normalized drift propagation and outside the cancellation regime, drift satisfies a conditional stability bound: |L̂ − L_true| ≤ (d/SHADES) × W_z + O(ε_mach). Within this regime drift is a certified bound, not merely an indicator. Cancellation is explicitly classified rather than silently degraded.

3.1 Encoding

Given x ∈ ℝ, x ≠ 0:
  σ  = sign(x)
  ℓ  = log₁₀|x|         [IEEE 754 Math.log10]
  z  = min{k ∈ ℤ : B_k > ℓ}   [zone boundaries {B_k}, half-open: B_{z-1} ≤ ℓ < B_z]
  W_z = B_z − B_{z−1}           [zone width in log₁₀ space]
  ι  = (ℓ − B_{z−1}) / W_z ∈ [0, 1)   [normalized intensity]
  s  = ⌊ι × 2047⌋                       [shade  - truncation, NOT round-to-nearest]
  δ₀ = ι × 2047 − s  ∈ [0, 1)          [initial quantization drift]

Zone boundary rule: values with ℓ = B_z exactly → zone z+1 (upper boundary excluded).
Rounding: floor (truncation toward zero). Systematic underestimate of up to 1/2047 shade.
Base case bound: |L̂ − L| ≤ (δ₀/SHADES) × W_z + O(ε_mach)  ✓ proven.

3.2 Addition (same sign, σ_a = σ_b)

ℓ_a = B_{z_a−1} + ι_a × W_{z_a}   [getLogValue()  - uses intensity, not shade]
ℓ_b = B_{z_b−1} + ι_b × W_{z_b}
ℓ_sum = ℓ_max + log₁₀(1 + 10^{ℓ_min − ℓ_max})   [stable log-sum-exp; avoids 10^ℓ overflow]

  Stable form: avoids materializing 10^â„“_a or 10^â„“_b directly.
  Safe for |ℓ| > 308. The 10^{ℓ_min − ℓ_max} term is always ≤ 1.

Drift update (zone-normalized  - satisfies conditional stability theorem):
  δ_result = δ₀_result + δ_a × (W_a / W_c) + δ_b × (W_b / W_c)

  W_a, W_b = zone widths of operands; W_c = zone width of result.
  Scaling by W/W_c converts operand drift into result-zone shade units.
  This is the only propagation rule that preserves the log-space error bound
  across zone boundaries. Fixed coefficients (e.g. × 0.01) are not correct.

3.3 Subtraction (σ_a ≠ σ_b) - Closed via log-diff-exp

ℓ_diff = ℓ_larger + log₁₀(1 − 10^{ℓ_smaller − ℓ_larger})   [log-diff-exp]

  Stable form: log₁₀(−expm1(Δ·ln10)) where Δ = ℓ_smaller − ℓ_larger ≤ 0.
  Math.expm1(x) = eˣ − 1, accurate near x ≈ 0 where naive 1 − exp(x) cancels.
  GeoNum NEVER projects to IEEE 754 for subtraction. Algebra is fully closed.

Conditioning: κ = 1 / |1 − 10^Δ|  (condition number of subtraction at this point)
  As ℓ_a → ℓ_b: Δ → 0, κ → ∞  (catastrophic cancellation  - fundamental, not fixable)

Drift amplification (derived from conditioning, not tuned):
  δ_result = (δ_a + δ_b) × κ   where κ = 1 / max(4ε_mach, |1 − exp(Δ·ln10)|)

  Floor 4ε_mach prevents division by zero at exact float equality.
  This is a structural floor derived from float64 representational limits, not a constant.

Near-cancellation (κ → ∞, result ≈ 0):
  Result classified as TRUST.CANCELLED with _cancelledLogMag = max(â„“_a, â„“_b).
  getUncertainty() returns the uncertainty at the cancelled scale, not zero.
  isUncertainZero() = true  - value is zero but not certain at this scale.

3.4 Multiply / Divide

â„“_product  = â„“_a + â„“_b   [exact in log space; IEEE 754 double precision]
ℓ_quotient = ℓ_a − ℓ_b

Drift update (zone-normalized  - satisfies conditional stability theorem):
  δ_result = δ₀_result + δ_a × (W_a / W_c) + δ_b × (W_b / W_c)

  Identical rule as addition. Log-space add/subtract of â„“ values is exact to
  IEEE 754 precision  - prior drift does not compound through the operation itself.
  Re-encoding the result introduces δ₀_result ∈ [0,1); zone-normalized operand
  drifts are added on top. This propagation is sufficient for the bound to hold
  inductively across arbitrary multiply/divide chains.

3.5 Zone Partition Contract

Valid zone config iff: B₀ < B₁ < B₂ < …  (strictly increasing)
  and each B_z = sup{ ℓ | ℓ ∈ zone z }   (each boundary is zone supremum)

  Enforced: setZoneConfig() throws RangeError on non-monotone input.
  Consequence: LUT binary search and linear scan produce identical results
  for all valid configs  - equivalence is guaranteed, not coincidental.
  Zone lookup: O(log n) binary search on monotone-filtered boundary table.

3.6 Trust Classification

getTrustLevel() → GeoNum.TRUST.*

  EXACT        drift = 0                   freshly encoded, no accumulated error
  PRECISE      0 < drift < 1              sub-shade; < 1 ULP-equivalent
  VALID        1 ≤ drift < SHADES/10      bounded; conditional stability theorem applies
  DEGRADED     SHADES/10 ≤ drift < SHADES/2   significant accumulated loss
  UNRELIABLE   drift ≥ SHADES/2           error spans half a zone width
  CANCELLED    isUncertainZero()          cancellation  - value meaningless at this scale

compareWeighted(other) → { cmp: -1|0|1, certain: bool }
  certain=false when uncertainty intervals overlap in log-space.
  Solvers should not treat DEGRADED/UNRELIABLE/CANCELLED results as convergence signals.

§ MPFR-256 Benchmark - GeoNum Rust/WASM

Both the JavaScript and Rust/WASM GeoNum implementations are validated independently against pre-computed MPFR-256 reference values (256-bit precision, 77 significant decimal digits). IEEE 754 cannot serve as a reference - each implementation is tested against ground-truth anchors, not against each other. Pass threshold: <0.5% relative error. Rust/WASM is the validated primary implementation.

Domain Coverage

Domain	Zone Config	Benchmark	Status
Theory	Lucas logarithmic	Hawking radiation	✓ 0.27% vs MPFR-256
GHD	sinh-compressed	Drude weight c→0	✓ live below
Quantum	eV-scale linear	-	zone config defined
Fluids	uniform grid	-	zone config defined
Plasma	frequency-aligned	-	zone config defined
Materials	lattice-symmetric	-	zone config defined

Angle for time-series: θ = 0° (fully open) θ = 9.43° θ = 18.87° θ = 28.30° θ = 37.73° θ = 47.17° (closure) Precision: GeoNum GeoNum Adv IEEE 754 Run Analysis

Rigid-Wall Baseline vs FSI Reference - Takizawa et al. 2014

Passage Geometry & Toroidal Integrals per θ

Tessellated Real-Space DFT via Adaptive Precision Zones

GPU-native Density Functional Theory using GeoNum zone tessellation instead of traditional plane-wave basis sets. Each zone is an independent compute unit - embarrassingly parallel on WebGPU compute shaders. Drift tracking provides automatic convergence diagnostics that traditional DFT codes lack.

Single Atom Diatomic Molecule Multi-Electron Atom Ab Initio MD Multi-Atom (Zone DFT) ORCA Input

Run DFT PES Scan GPU Poisson

DFT Engine - Validated Capabilities

HPC Lab tier required. Contact sales.

HPC Lab tier required. Contact sales.

HPC Lab tier required. Contact sales.

HPC Lab tier required. Contact sales.

HPC Lab tier required. Contact sales.

Scope: this is a surrogate model, not a replication or falsification of any quantum-hardware benchmark. It evolves a synthetic structured circuit under a white-noise model - not real device circuit artifacts (QASM/QPY) and not returned hardware counts. It illustrates the scoring mechanism (XEB/HOG, same-layout controls, GeoNum substrate precision). An actual replication requires ingesting the real circuit artifacts and the returned hardware counts and scoring those through the same pipeline.

HPC Lab tier required. Contact sales.

What this is: paste anyone's published OpenQASM 2/3 circuit (Qiskit, Cirq, IBM, IonQ all export it). The 2-qubit interaction graph is split into independent patches; every patch up to 24 qubits is evolved as an exact independent statevector, so a patch-separable circuit is full-precision validated up to 128 qubits as the tensor product of its patches - never 2^128. If you also paste the author's measured bitstring counts, we run patch-XEB against our independent ideal (the same per-patch cross-entropy used for RCS), with an automatic bit-order probe. A patch wider than 24 qubits is reported BEYOND-REACH with no fabricated number - that genuinely needs a cluster, which is exactly where a published ideal is itself precision-contested. Download a self-describing, replayable bundle so anyone can reproduce the verdict without a cluster.

HPC Lab tier required. Contact sales.

What this is: the classical half of quantum-centric supercomputing (the IBM / Cleveland Clinic / RIKEN SQD method), run entirely in your browser with no cluster. Ingest a published active space (FCIDUMP: h_pq, (pq|rs) chemist notation, E_nuc, NORB, NELEC, MS2) plus sampled electronic configurations - either real published quantum samples or a clearly-labelled classical CISD surrogate. We run self-consistent configuration recovery (S-CORE), project the Hamiltonian onto the sampled-determinant subspace by Slater-Condon rules, and find the ground state by Davidson. This reproduces / audits the active-space SQD energy on classical hardware. The molecular environment (the 12,635-atom protein) is not computed here - a supplied classical embedding term is reported separately and never claimed as ours. Subspace dimensions past the browser cap are reported BEYOND-REACH with no fabricated energy: that regime genuinely needs the supercomputer, which is exactly the SQD premise. The Slater-Condon engine is validated every run against closed-form references (Hubbard dimer, non-interacting limit, H2/STO-3G FCI).

HPC Lab tier required. Contact sales.

What this is: a 2D extension of Nudged Elastic Band. Instead of a single minimum-energy path, NEM relaxes a 2D membrane of images on the potential energy surface to expose multiple MEPs, transition states, second-order saddles, shared intermediates, and post-TS bifurcations in one pass. Analytic benchmark surfaces (Mueller-Brown, Wolfe-Quapp, Cerjan-Miller, LEPS) run on GPU in seconds with full GeoNum drift tracking per node. DFT-backed surface uses the existing diatomic engine with a per-node 30 s wall cap.

HPC Lab tier required. Contact sales.

References: Mueller & Brown, Theor. Chim. Acta 53, 75 (1979). Quapp, J. Theor. Comput. Chem. 2, 385 (2003) (Wolfe-Quapp surface). Cerjan & Miller, J. Chem. Phys. 75, 2800 (1981). Sato, J. Chem. Phys. 23, 592 (1955) (LEPS). Henkelman & Jonsson, J. Chem. Phys. 113, 9978 (2000) (NEB / nudging projection). NEM formalism: Muechler et al. 2025 (motivating paper for this tab).

Solves the Euler-Lagrange boundary-value problem d/dx(dL/dq') - dL/dq = 0 by minimizing the discrete action on a uniform grid. Same regularized-Newton substrate as the NEM tab. Per-node drift comes from Rust GeoNum (wasm geonum_evaluate). Each Lagrangian validates against its closed-form analytic solution (cycloid, sinusoid, great circle).

HPC Lab tier required. Contact sales.

References: Bernoulli J., Acta Eruditorum (1696) brachistochrone problem, solution = cycloid. Riemann, Habilitationsschrift (1854) geodesics. Standard harmonic-oscillator BVP per Landau & Lifshitz, Mechanics §5. Discrete-action variational principle per Marsden & West, Acta Numerica 10, 357 (2001).

A single cavity photon mode Zeeman-couples to the collective magnetization of a transverse-field Ising model across its quantum critical point. When the coupling is parallel to the order parameter (n = x) the ferromagnetic and superradiant phases cooperate and the critical coupling g_c vanishes at the QCP; when orthogonal (n = y) they compete, giving a tricritical point and first-order transitions. Reproduces the phase diagram, polariton spectrum, intrinsic squeezing and quantum Fisher information of Sur, Wang, Mahankali, Paschen & Si, Nat. Commun. 17, 4404 (2026). Every reported quantity carries a Rust GeoNum drift / trust tag. The TFIM lattice coordination number z is a geometric property of the chosen mesh (the GMDBS pentagon mesh has z = 7).

HPC Lab tier required. Contact sales.

Reference: Sur, Wang, Mahankali, Paschen & Si, "Amplified response of cavity-coupled quantum-critical systems", Nature Communications 17, 4404 (2026), doi:10.1038/s41467-026-73112-1. Phase boundary Eq. 8; squeezing effective Hamiltonian Supplementary Note 3 (Eq. S31); QFI Eq. 3. Polariton spectrum derived from the two-mode Hopfield diagonalization, normalized to reproduce the paper's Eq. 8 phase boundary and Dicke limit. Squeezing reported from the soft-polariton ratio (perfect intrinsic squeezing as E_- -> 0).

A fundamental quasinormal ringdown (l = m = 2, n = 0) for a Kerr black hole is laid down on an extraction sphere as a Weyl scalar psi4(theta, phi, t), decomposed onto spin-weighted spherical harmonics -2Y_lm (Wigner-d form), converted to strain h(t) by fixed-frequency double integration, and fit for the complex quasinormal frequency Mω, which is compared to the published Kerr value. Each extracted mode carries a GeoNum drift / trust tag: the same operation that computes the amplitude reports whether it is physical signal or numerical cancellation. Scope: this validates the extraction and analysis chain on a reference psi4; wiring it to a live BSSN-evolved Weyl tetrad is the next step.

HPC Lab tier required. Contact sales.

References: Goldberg et al., J. Math. Phys. 8, 2155 (1967) - spin-weighted harmonics; Leaver, Proc. R. Soc. Lond. A 402, 285 (1985) - Schwarzschild QNM; Berti, Cardoso & Will, Phys. Rev. D 73, 064030 (2006) - Kerr QNM table; Reisswig & Pollney, Class. Quantum Grav. 28, 195015 (2011) - fixed-frequency psi4 -> h integration. Quasinormal frequencies in geometric units Mω (G = c = M = 1).

A Fibonacci-chain quasicrystal (off-diagonal tight-binding, two hoppings t_A, t_B in the Fibonacci sequence A→AB, B→A) is solved by the linear-scaling Kernel Polynomial Method: Chebyshev moments built from one sparse matrix-vector each, Jackson-damped, reconstructed into the density of states - no diagonalization, O(N) so it scales to millions of sites. The spectrum is a Cantor set whose gaps carry integer topological labels m via the gap-labeling theorem, IDOS = {mτ} mod 1 (τ the golden ratio). Same problem as the Aalto tensor-network result (Antao, Sun, Fumega & Lado, Phys. Rev. Lett. 136, 2026); here with a transparent method and a GeoNum drift band on the Chebyshev reconstruction. Scope: 1D Fibonacci chain on the CPU; GPU sparse mat-vec for the million-site regime and 2D Penrose / Ammann-Beenker tilings are the next steps.

HPC Lab tier required. Contact sales.

References: Weisse, Wellein, Alvermann & Fehske, Rev. Mod. Phys. 78, 275 (2006) - Kernel Polynomial Method; Kohmoto, Kadanoff & Tang, Phys. Rev. Lett. 50, 1870 (1983) and Bellissard et al. (1989) - Fibonacci gap labeling; motivated by Antao, Sun, Fumega & Lado, Phys. Rev. Lett. 136 (2026), Aalto University. Exact integrated DOS computed by Sturm sequence (negative-pivot count); gap labels by best-fit integer m in {mτ} mod 1.

A 2D discrete nonlinear Schrodinger lattice (light in a nonlinear waveguide array): i ∂_tc = κ∑_neighborsc + χ|c|²c + Uδ_n,mc. From a near-uniform input, modulational instability drives the field toward self-trapped (localized) states; a refractive-index detuning U on the diagonal (a Hubbard-like interaction) opens intermediate breathing and chaotic-like regimes. The conserved norm carries a GeoNum reliability band - the substrate's drift horizon for long nonlinear integration. Reproduces Bezerra & Dias, Nonlinear Dynamics 114:731 (2026).

HPC Lab tier required. Contact sales.

Anchors: CW dispersion ω = 2κ(cos Q_x+cos Q_y)+χ|A|²; modulational-instability gain g(Q)=√(-f(f+2χ|A|²)), f=2κ(cos Q_x+cos Q_y-2); norm + Hamiltonian conservation; exact two-site self-trapping threshold χN/κ = 4 (Kenkre-Campbell / Eilbeck-Lomdahl-Scott). Reproduces Bezerra & Dias, Nonlinear Dynamics 114:731 (2026).