Claim-to-Artifact Manifest
All paths below are repo-root-relative.
- Long-form technical report:
docs/paper_combined/main.tex - NeurIPS 9-page main-track submission:
docs/paper_neurips/main.tex(when present; seedocs/paper_neurips/for the packaged submission)
Numeric headline audit: docs/paper_combined/CLAIM_INVENTORY.md.
Section 5: Does Hybridization Help?
| Claim | Artifact (repo-root-relative) | Regeneration |
|---|---|---|
| Hybrid vs PINN-only (Heat1D d=-8.3, Wave2D d=-3.6) | experiments/results/rq3_10seed/results.json, experiments/results/rq4_10seed/results.json |
experiments/scripts/rq3_model_comparison.py, experiments/scripts/rq4_wave2d_model_comparison.py |
| KdV hybrid (d=-0.38), AllenCahn (d=-3.12) | experiments/results/rq1e_kdv_10seed/results.json, experiments/results/rq1f_allencahn_10seed/results.json |
experiments/scripts/rq1e_kdv_hybrid.py, experiments/scripts/rq1f_allencahn_hybrid.py |
| PINO comparison (2 PDEs) | experiments/results/rq_pino_heat1d_10seed/results.json, experiments/results/rq_pino_advdiff/results.json |
experiments/scripts/rq_pino_comparison.py, experiments/scripts/rq_pino_advdiff.py |
| OOD evaluation (coefficient, domain, IC/domain shifts) | experiments/results/rq6_ood/results.json, experiments/results/rq6_ood_heat1d/results.json, experiments/results/rq6_ood_advdiff/results.json, experiments/results/rq6b_ood_domain/results.json, experiments/results/rq6c_ood_ic_shift/results.json |
experiments/scripts/rq6_ood_*.py |
| HyPINO official native-suite smoke (context only; not same-PDE head-to-head) | experiments/results/paper_combined/hypino_official_eval/results.txt |
experiments/scripts/paper_combined/evaluate_hypino_official.py |
| HyPINO zero-shot Heat1D adapter smoke (compatibility only; not same-budget baseline) | experiments/results/paper_combined/hypino_heat1d_adapter/results.json |
experiments/scripts/paper_combined/evaluate_hypino_heat1d_adapter.py |
| HyPINO target-PINN adaptation probe (Heat1D + AdvDiff1D, 500 Adam steps, 3 seeds) | experiments/results/paper_combined/hypino_adaptation_2pde/results.json |
experiments/scripts/paper_combined/finetune_hypino_heat1d.py, experiments/scripts/paper_combined/finetune_hypino_advdiff1d.py, experiments/scripts/paper_combined/aggregate_external_probes.py |
| PINNacle executable subset smoke (Burgers1D/Wave1D/Poisson1D/Helmholtz2D/Heat2D-Multiscale, 200 iterations each) | experiments/results/paper_combined/pinnacle_subset_5task/results.json |
experiments/scripts/paper_combined/run_pinnacle_subset.py, experiments/scripts/paper_combined/aggregate_external_probes.py |
| No-dominance (13 PDEs) | experiments/results/paper_combined/routing_evaluation/results.json |
experiments/scripts/paper_a_routing/run_policy_suite.py |
Section 6: Does Causal Loss Help?
| Claim | Artifact (repo-root-relative) | Regeneration |
|---|---|---|
| 2x2 stage ablation (9 PDEs, 360 runs) | experiments/results/paper_combined/stage_ablation/*.json |
experiments/scripts/paper_c_causal/stage_ablation.py |
| Statistical tests (BH-FDR, 27 tests) | experiments/results/paper_combined/statistical_tests.json |
experiments/scripts/statistical_analysis.py |
| Gradient flow (cos~1.0, 4 PDEs) | experiments/results/paper_c/gradient_flow_*/gradient_flow_results.json |
experiments/scripts/paper_c_causal/gradient_flow_analysis.py |
| Temporal error (ratio 0.98-672x) | experiments/results/paper_c/temporal_error_analysis/results.json |
experiments/scripts/paper_c_causal/temporal_error_analysis.py |
| Causal weight collapse | experiments/results/paper_combined/adaptive_epsilon/results.json |
experiments/scripts/paper_combined/test_adaptive_epsilon.py |
| Wang-style single-stage causal comparison (4 PDEs) | experiments/results/paper_combined/wang_comparison/*_results.json |
experiments/scripts/paper_c_causal/wang_comparison.py |
| Adaptive epsilon mitigation | experiments/results/paper_combined/adaptive_epsilon/results.json |
experiments/scripts/paper_combined/test_adaptive_epsilon.py |
Section 7: Physics Weight Tradeoff
| Claim | Artifact (repo-root-relative) | Regeneration |
|---|---|---|
| Wave2D lambda sweep (5 values) | experiments/results/rq5_wave2d_lambda_sweep/results.json |
experiments/scripts/rq5_wave2d_lambda_sweep.py |
| Burgers1D lambda sweep | experiments/results/rq2b/results.json |
experiments/scripts/rq2b_burgers_lambda_sweep.py |
Section 8: Policy Selection from Early Dynamics
| Claim | Artifact (repo-root-relative) | Regeneration |
|---|---|---|
| Routing evaluation (13 PDEs, 450 runs) | experiments/results/paper_combined/routing_evaluation/results.json |
experiments/scripts/paper_a_routing/evaluate_routing_full.py |
| Scaling analysis (5→9→13 PDEs) | experiments/results/paper_combined/routing_evaluation/scaling_ablation.json |
experiments/scripts/paper_a_routing/scaling_and_ablation.py |
| Feature ablation (nested CV) | experiments/results/paper_combined/routing_evaluation/scaling_ablation.json |
Same script (Part 2) |
| Diagnostic bound (33/52 = 63.5% overall; 33/39 = 84.6% for k ∈ {10,50,100}; universal failure at k=20; fails on Burgers1D and Burgers1D-lownu at every defined k) | experiments/results/paper_combined/routing_evaluation/regret_bound_validation.json |
experiments/scripts/paper_combined/validate_regret_bound.py |
| Meta-router (GBR+PDE and RF+PDE-top5: 84.6%, 11/13 held-out PDEs; full RF+PDE: 61.5%) | experiments/results/paper_combined/meta_router/results.json |
experiments/scripts/paper_combined/evaluate_meta_router.py |
| Held-out generalization (3 PDEs) | experiments/results/paper_combined/routing_evaluation/holdout.json |
experiments/scripts/paper_a_routing/holdout_validation.py |
| Family-macro accuracy (LOPO aggregated by 8 families): physics-final 81.2% / GBR 68.8% / RF 54.2% / Ridge 45.8% | experiments/results/paper_combined/routing_evaluation/results.json |
experiments/scripts/paper_a_routing/family_holdout_analysis.py |
| Bandit comparison (SH, LinUCB) | experiments/results/paper_a/bandit_evaluation/results.json |
experiments/scripts/paper_a_routing/evaluate_bandits.py |
| Accuracy-regret dissociation (77x) | experiments/results/paper_combined/routing_evaluation/results.json + experiments/results/rq_early_predictor_l2/results.json |
experiments/scripts/paper_a_routing/evaluate_routing_full.py |
Figures
All figure sources live under docs/paper_combined/figures/ (repo-root-relative).
| Figure | Regeneration |
|---|---|
fig_no_dominance.pdf |
experiments/scripts/paper_a_routing/generate_figures.py |
fig_routing_accuracy.pdf |
Same |
fig_feature_importance.pdf |
Same |
fig_budget_savings.pdf |
Same |
fig_scaling.pdf |
experiments/scripts/paper_a_routing/scaling_and_ablation.py |
fig_feature_ablation.pdf |
Same |
fig_dissociation.pdf |
experiments/scripts/paper_a_routing/generate_dissociation_fig.py |
fig_ntk_proxy.pdf |
experiments/scripts/paper_a_routing/ntk_proxy_analysis.py |
fig_regret_bound.pdf |
experiments/scripts/paper_combined/validate_regret_bound.py |
fig_stage_ablation_heatmap.pdf |
experiments/scripts/paper_c_causal/generate_figures.py |
fig_pretrain_effect.pdf |
Same |
fig_interaction_plot.pdf |
Same |
fig_runtime_tradeoff.pdf |
Same |
fig_gradient_alignment.pdf |
experiments/scripts/paper_c_causal/gradient_flow_analysis.py |
fig_gradient_combined.pdf |
experiments/scripts/paper_c_causal/generate_combined_gradient_fig.py |
fig_causal_weights.pdf |
experiments/scripts/paper_c_causal/causal_weight_evolution.py |
fig_temporal_error.pdf |
experiments/scripts/paper_c_causal/temporal_error_analysis.py |
fig_epsilon_trajectories.pdf |
experiments/scripts/paper_combined/test_adaptive_epsilon.py |
fig_weight_collapse.pdf |
Same |
Table generation provenance (hand-curated vs auto-generated)
The paper_combined manuscript contains roughly twenty tables. Their provenance with respect to the released JSON artifacts is as follows.
- Auto-generated (from JSON via
experiments/scripts/generate_paper_tables.pyorexperiments/scripts/statistical_analysis.py):tab:stats-corestatistics rows,tab:causal-ablationper-PDE effect rows, and the raw fragments inexperiments/analysis/paper_tables.texandexperiments/analysis/statistical_tests.md. - Hand-curated from JSON numeric values (values copy-pasted from the
corresponding
results.json; verified inCLAIM_INVENTORY.mdand byscripts/verify_claims.py):tab:routing,tab:meta_router,tab:feature-importance,tab:scaling,tab:bestpolicy,tab:lopo-family,tab:adaptive-epsilon, OOD taxonomy table, Wang single-stage comparison table, HyPINO adaptation table, and the PINNacle subset table.
Hand-curated tables intentionally remain hand-curated for this submission: the
values match the inventory, each row cites a specific artifact path, and a
combined-paper-wide table generator is deferred (out of scope for the reject-risk
reduction pass). Any future numeric change must be reflected simultaneously in
the JSON artifact, the paper table, and CLAIM_INVENTORY.md.
Notes on scope (per REVISION.md)
- All experiments are in the synthetic exact-solution regime; Stage 1 supervised data comes from analytical solutions. No claim is made about noisy, sparse, or real-measurement regimes.
- External baselines (HyPINO, PINNacle, Wang-style causal, adaptive weighting) are contextual positioning except for the internal PINO comparison (Heat1D + AdvDiff1D) and the Wang-style 4-PDE probe, which are direct same-pipeline probes but not broad head-to-head reproductions.
- The routing suite is 13 PDE configurations across 8 families; results are conditioned on that composition.
pinn_onlydenotes the PINN backbone under the same supervised-data access as the hybrid, not a standard data-free PINN. It isolates architecture, not data regime.