hotSpring Phase B Evidence

Date: February 27, 2026
Purpose: Document how hotSpring validation provides quantitative evidence for the constrained evolution thesis
Context: BarraCuda has now been validated through all phases: A (Python control, 86/86), B (nuclear EOS), C (GPU MD, 9/9), D (native f64 builtins, N-scaling), E (paper-parity long run, 80k steps), plus lattice QCD (6 papers, 32⁴ production), spectral theory (4 papers, 45/45), transport coefficients (13/13 Green-Kubo), screened Coulomb (23/23), and NPU integration (Exp 020-022, live AKD1000). Total: 22 papers, ~700 checks, 39/39 suites.

1. The Constrained Evolution Claim

The gen3 thesis states: selective pressure from constrained environments (Rust, Pure Rust, capability-based design) drives convergent evolution — the same mathematical operations emerge because the same underlying math is needed, regardless of whether the driving workload is ML, FHE, or physics.

hotSpring provides the first quantitative test of this claim.

2. Evidence: ML-Evolved Math Serves Physics

The BarraCuda Scientific Compute Gaps (Feb 7) predicted that “ML-driven evolution covered ~60-70% of what scientific computing needs.” hotSpring’s L2 validation measured the actual coverage:

Functions that ML evolution produced and physics validated:

Measured coverage: 11/11 core functions worked for physics. The 60-70% estimate was conservative — for this particular workload, 100% of needed functions existed (though some needed fixes).

Functions that physics revealed need evolution:

Key finding: The gaps are at the precision/specialization boundary, not the mathematical foundation. Physics doesn’t need different math — it needs the same math done more carefully.

3. Quantitative Evidence of Evolution

3.1 The 1,764x Improvement Through Iterative Evolution

Cycle 1: hotSpring reports missing physics           → 28,450 → 92     (309x)
Cycle 2: hotSpring identifies numerical precision     → 92 → 25         (3.7x)
Cycle 3: ToadStool delivers brent + eigh_f64         → 25 → 16.11      (1.5x)
Cycle 4: hotSpring validates, identifies GPU targets  → ready for Titan V

Each cycle follows the constrained evolution pattern:

1. Selection pressure (hotSpring validation reveals a gap)
2. Handoff (wateringHole documents the gap with code locations)
3. Evolution (ToadStool team implements the fix)
4. Validation (hotSpring confirms improvement)

3.2 Convergent Evolution with SciPy

The most striking evidence: BarraCuda’s eigh_f64 (Jacobi algorithm) evolved from ML linear algebra needs, yet when validated against numpy.linalg.eigh (LAPACK) on HFB Hamiltonians, it produces matching eigenvalues to machine precision. The algorithm is different (Jacobi vs Householder+QR), but the mathematical convergence is identical.

Similarly, brent was implemented for general optimization, but when used for BCS chemical potential (a physics-specific application), it matches scipy.optimize.brentq to 10^-10 precision.

This is convergent evolution: different selective pressures (ML vs physics) produce functionally equivalent tools.

3.3 Zero-Dependency Achievement

The removal of nalgebra (the last external dependency) from hotSpring’s HFB solver demonstrates the thesis that a sufficiently evolved Pure Rust library can replace the scientific Python stack:

Before: hotSpring → nalgebra → LAPACK → Fortran
After:  hotSpring → barracuda::eigh_f64 (Pure Rust, zero deps)

This mirrors the biological analogy: a specialized organism (BarraCuda) eventually internalizes all capabilities needed for its ecological niche.

4. Phase C Evidence: GPU Molecular Dynamics as Convergent Evolution

Phase C provides the strongest constrained evolution evidence yet. The f64 WGSL shaders built for nuclear EOS (Phase B) were extended to run full Yukawa OCP molecular dynamics entirely on a consumer GPU (RTX 4070, $350).

4.1 ML Math Serves MD Physics (Convergent Reuse)

The GPU MD pipeline uses the same mathematical substrate that ML workloads evolved:

The same math library that trains neural networks now runs plasma physics simulations. This is not a theoretical prediction — it is measured.

4.2 Results

4.3 What Phase C Proves About Constrained Evolution

1. ML-evolved math is physics-complete for MD. Every mathematical operation needed for Yukawa OCP simulation already existed in BarraCuda from ML/FHE development. No new math functions were required — only new compositions of existing functions.

2. The GPU substrate is the same for ML and physics. The same SHADER_F64 capability that accelerated nuclear EOS evaluation (Phase B) runs molecular dynamics force kernels. One GPU, one shader language, two domains.

3. Consumer hardware is sufficient. A $350 RTX 4070 runs production plasma physics at 74-120 steps/s with exact energy conservation. This is the sovereignty thesis in action — no HPC cluster, no CUDA, no institutional access required.

5. Performance Comparison as Evolution Metric

The evolved system outperforms the incumbent on every dimension measured except L2 accuracy, where the gap is sampling strategy (SparsitySampler), not physics or compute. The lattice QCD and NPU capabilities have no Python-stack equivalent — they represent evolution beyond the incumbent’s niche.

6. Predictions for Next Evolution Cycle

Based on the constrained evolution framework, the next selective pressures are:

Already confirmed (Phase C validated):

1. PBC integration: Physics required periodic boundary conditions. hotSpring built them into force/drift shaders. Prediction confirmed.
2. Symplectic integrator: Physics required energy-conserving time integration. hotSpring built split Velocity-Verlet. Prediction confirmed.
3. Thermostat: Physics required temperature control. hotSpring built Berendsen thermostat. Prediction confirmed.
4. GPU-native observables: Physics required RDF, VACF, SSF computed from GPU snapshots. hotSpring built the pipeline. Prediction confirmed.

Confirmed (Phases D-F, Feb 2026):

5. Large-N scaling: N=10,000 MD in 5.3 min on RTX 3090 with cell-list optimization (4.1× faster). Confirmed.
6. Paper-parity long run: 9/9 PP Yukawa cases × 80,000 steps, $0.044 total. 0.000% drift. Confirmed.
7. 3D FFT pipeline: ToadStool Fft1DF64/Fft3DF64 achieved, roundtrip error 1e-10. FFT gap closed. Confirmed.
8. DF64 Core Streaming: Discovery — 3.24 TFLOPS at 14-digit precision on RTX 3090 FP32 cores (9.9× native f64 throughput). Lattice QCD trajectories 2× faster. Novel contribution.
9. Lattice QCD production: 32⁴ pure gauge SU(3) β-scan, deconfinement at β_c=5.69, 13.6 hours, $0.58. Confirmed.
10. NPU adaptive steering: Live AKD1000 ESN inference during HMC (Exp 022). 63% thermalization savings, cross-run learning. Novel contribution.

Remaining:

• Batched eigendecomposition: L2 HFB needs 52 simultaneous Hamiltonian diagonalizations on GPU.
• 2D numerical methods: Deformed nuclei need 2D grids.
• WDM extension: Partial ionization, quantum corrections (Murillo Reproduction Plan Tier 4).

Each is a testable prediction. hotSpring validates as hardware and capabilities arrive.

7. Data Sources

See also:

• hotSpring Results — thesis chapter on hotSpring validation
• Quantitative Evidence — cross-spring measured outcomes
• hotSpring Validation Summary — live validation status and binaries