Dataset and Reproduction Package for A Guarded Trainable Exponent Operator: Structure-Matched Extrapolation, Loss Targets, and Numerical Boundaries

EnvironmentPython 3.10 (environment.yml). CPU-only, Ubuntu 22.04. Seeds 0-23 main, 20001-20002 pilot disjoint. conda env create -f environment.yml Dataset layout/synthetic/: 1728 cells (n=32,64,256; sigma=0,0.01,0.05,0.1,0.2,0.5; 24 clusters) x (128 diag+256 ID+256 OOD+128 near-zero)=10368 primary+3096 diagnostic. ID log x∈[-1,1], OOD |log x|∈[1.2,2.2], near-zero x1∈[1e-10,1e-6] log-uniform./real/yacht/: 308 rows, 22 hulls (UCI). Feature2 non-positive → Yacht5 (positive-only) and Yacht6 (raw linear term)./real/chwirut2/: 54 rows, 22 distances (NIST)./checkpoints/: 10368 FP64 L-BFGS-B checkpoints. Main grid (Table3, Fig1-4)python src/generate_synthetic.py --seeds 0-23python src/train.py --config configs/power_grid.yaml --loss log_mse --optimizer lbfgsb --max_iter 300 --ftol 1e-12 --gtol 1e-8python src/evaluate.py --split id ood nearzero --metric nmae rmsenMAE=mean(|ŷ-y|/(|y|+1)), RMSE=sqrt(mean((ŷ-y)^2)). Guarded Eq1: y=exp(clip[b+Σqj log max(xj,1e-8),-30,30]). Lemma1 & identifiability (Fig2)python src/covariance_calibration.py --corr 0 0.95 0.99999 1 Verifies Cov=σ²(DᵀD)⁻¹ and shift exp(-σ²/2): 0.0049 at σ=0.1, 0.1175 at 0.5. Fixed & misspec (Table4-6)python src/train.py --config configs/fixed_q.yaml --q 0.25 1.0 --fit_only_bpython src/train.py --config configs/nonpower.yamlpython src/train.py --config configs/budget_1000.yaml --max_iter 1000 Real validation (Table7, Fig5, FigS2)python src/train_real.py --dataset yacht --outer leave-one-hull-out --tracks yacht5 yacht6python src/train_real.py --dataset chwirut2 --outer leave-one-distance-out GINN-LP budget (Table8)Frozen commit 165f32b, 32 synthetic (n=64,σ=0.1,8 seeds)+22 Yacht5 folds, ≤4 terms, 2 restarts, 200 epochs/cand (100 reg+100 unreg, L1=L2=1e-4, total 1600). Precision & footprint (Table9, Fig6, TableS1-S2)python src/precision_audit.py --storage fp64 fp32 fp16 --compute fp32 → 93312 summaries, 100% finite.python src/benchmark.py --pinned_cpu 1 --warmup 30 --blocks 11x100python src/ablation.py --epsilon 1e-12 1e-8 1e-6 --L 10 20 30 50 (Table S1)python src/train.py --config configs/mlp64_adam.yaml --epochs 2000 (Table S2) Reference impldef guarded_power(x,q,b,eps=1e-8,L=30): z=b+log(maximum(x,eps))@q; return exp(clip(z,-L,L)) All tables/figures from results/*.json via src/figures.py and src/tables.py.