Upload scripts/generate_paper_figures.py with huggingface_hub

scripts/generate_paper_figures.py

@@ -2,8 +2,9 @@
 ⚡ ULTRON Systems // 100% Deterministic Empirical Figure Generator (300 DPI)
 Zero synthetic data. Reads directly from:
 - logs/pretraining_telemetry_verified.json (855 real logged pre-training steps)
-- logs/oos_evaluation_results.json (95,040-window 27-shard OOS benchmark)
-- logs/tier1_verified_audit.json (Multi-epoch, Frobenius drift, 7-horizon ablation, causal backtest)
+- logs/defensible_tier1_audit.json (Stratified 15-ticker contiguous OOS & Pre-2024 Stress audit,
+  exact global cosine similarities, 1m return-magnitude buckets, paired bootstrap CIs,
+  mean-ablation + random-channel controls, and non-overlapping state-turnover backtests)
 """
 import os
 import json
@@ -17,21 +18,20 @@ plt.style.use('seaborn-v0_8-whitegrid' if 'seaborn-v0_8-whitegrid' in plt.style.
 mpl.rcParams['font.family'] = 'serif'
 mpl.rcParams['font.size'] = 9.5
 mpl.rcParams['axes.labelsize'] = 10
-mpl.rcParams['axes.titlesize'] = 11
+mpl.rcParams['axes.titlesize'] = 10.5
 mpl.rcParams['xtick.labelsize'] = 8.5
 mpl.rcParams['ytick.labelsize'] = 8.5
-mpl.rcParams['legend.fontsize'] = 8
+mpl.rcParams['legend.fontsize'] = 7.8
 mpl.rcParams['figure.titlesize'] = 12
 
 with open("logs/pretraining_telemetry_verified.json") as f:
     pre_data = json.load(f)
 
-with open("logs/oos_evaluation_results.json") as f:
-    oos_27 = json.load(f)
-
-with open("logs/tier1_verified_audit.json") as f:
+with open("logs/defensible_tier1_audit.json") as f:
     audit = json.load(f)
 
+horizons = ["1m", "3m", "5m", "15m", "30m", "60m", "1d"]
+
 # =========================================================================
 # Figure 1: REAL Empirical Pre-Training Telemetry (855 Logged Steps)
 # =========================================================================
@@ -41,7 +41,6 @@ steps = np.array([r["step"] for r in records])
 t_loss = np.array([r["total_loss"] for r in records])
 d_acc = np.array([r["dir_acc_total"] for r in records])
 
-# Compute strictly causal 25-point moving average for trend visibility alongside raw points
 def causal_sma(arr, w=25):
     out = np.zeros_like(arr, dtype=np.float64)
     for i in range(len(arr)):
@@ -80,60 +79,53 @@ fig.savefig("papers/figures/fig1_pretraining_curves.png", dpi=300, bbox_inches='
 plt.close()
 
 # =========================================================================
-# Figure 2: 27-Shard OOS Benchmark + Epoch 1/2/3 Progression vs Baselines
+# Figure 2: 1m Base-Rate / Flat-Bar Decomposition, Return Buckets & Baselines
 # =========================================================================
-print("Rendering Figure 2 from real OOS benchmark & Epoch progression...")
-horizons = ["1m", "3m", "5m", "15m", "30m", "60m", "1d"]
-acc_2024 = [oos_27["static_oos_zero_shot"]["2024"]["metrics"]["per_horizon"][h]["dir_acc"] for h in horizons]
-acc_2025 = [oos_27["static_oos_zero_shot"]["2025"]["metrics"]["per_horizon"][h]["dir_acc"] for h in horizons]
-acc_2026 = [oos_27["static_oos_zero_shot"]["2026"]["metrics"]["per_horizon"][h]["dir_acc"] for h in horizons]
-
+print("Rendering Figure 2 from 1m return-magnitude buckets & multi-horizon ROC-AUC...")
 fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(11, 4.0), dpi=300)
 
+# Panel (a): All-Bar ROC-AUC vs Strict Non-Zero ROC-AUC across Horizons (2024 OOS, N=8,044)
+ep3_24 = audit["multi_checkpoint_oos_and_stress_audit"]["epoch3"]["2024"]["per_horizon"]
+ep3_26 = audit["multi_checkpoint_oos_and_stress_audit"]["epoch3"]["2026"]["per_horizon"]
+auc_all_24 = [ep3_24[h]["roc_auc"] for h in horizons]
+auc_nz_24 = [ep3_24[h]["nonzero_roc_auc"] for h in horizons]
+auc_nz_26 = [ep3_26[h]["nonzero_roc_auc"] for h in horizons]
+
 x = np.arange(len(horizons))
-w = 0.25
-ax1.bar(x - w, acc_2024, w, label='2024 OOS (42,240 win)', color='#0284C7', edgecolor='black', lw=0.5)
-ax1.bar(x, acc_2025, w, label='2025 OOS (42,240 win)', color='#0D9488', edgecolor='black', lw=0.5)
-ax1.bar(x + w, acc_2026, w, label='2026 OOS (10,560 win)', color='#D97706', edgecolor='black', lw=0.5)
-ax1.axhline(50.0, color='#E11D48', linestyle='--', lw=1.2, label='Random Walk Null (50.0%)')
+w = 0.26
+ax1.bar(x - w, auc_all_24, w, label='2024 All-Bar ROC-AUC', color='#0284C7', edgecolor='black', lw=0.5)
+ax1.bar(x, auc_nz_24, w, label='2024 Non-Zero (y≠0) ROC-AUC', color='#0D9488', edgecolor='black', lw=0.5)
+ax1.bar(x + w, auc_nz_26, w, label='2026 Non-Zero (y≠0) ROC-AUC', color='#D97706', edgecolor='black', lw=0.5)
+ax1.axhline(0.5000, color='#E11D48', linestyle='--', lw=1.2, label='Random Guess / Flat Heuristic (0.5000)')
 ax1.set_xticks(x)
 ax1.set_xticklabels(horizons)
-ax1.set_ylim(46.0, 56.0)
-ax1.set_ylabel('Directional Accuracy (%)')
-ax1.set_title('(a) 95,040-Window Zero-Shot OOS Accuracy by Horizon')
+ax1.set_ylim(0.480, 0.545)
+ax1.set_ylabel('Threshold-Independent ROC-AUC')
+ax1.set_title('(a) All-Bar vs. Strict Non-Zero (y≠0) ROC-AUC by Horizon')
 ax1.legend(loc='upper right', frameon=True)
 ax1.grid(axis='y', alpha=0.3)
 
-# Panel (b): Epoch 1 vs Epoch 2 vs Epoch 3 vs Baselines (2024 OOS Cohort)
-models_lbl = ["Always\nLong", "1-Bar\nMom", "EMA\n12/26", "5-Bar\nRev", "Conformer\nEpoch 1", "Conformer\nEpoch 2", "Conformer\nEpoch 3"]
-raw_vals = [
-    audit["quantitative_baselines_oos"]["2024"]["Always_Long_Drift"]["overall_raw_dir_acc"],
-    audit["quantitative_baselines_oos"]["2024"]["Momentum_1Bar"]["overall_raw_dir_acc"],
-    audit["quantitative_baselines_oos"]["2024"]["EMA_12_26_Crossover"]["overall_raw_dir_acc"],
-    audit["quantitative_baselines_oos"]["2024"]["MeanReversion_5Bar"]["overall_raw_dir_acc"],
-    audit["multi_checkpoint_oos_audit"]["epoch1"]["2024"]["overall_raw_dir_acc"],
-    audit["multi_checkpoint_oos_audit"]["epoch2"]["2024"]["overall_raw_dir_acc"],
-    audit["multi_checkpoint_oos_audit"]["epoch3"]["2024"]["overall_raw_dir_acc"]
-]
-nz_vals = [
-    audit["quantitative_baselines_oos"]["2024"]["Always_Long_Drift"]["overall_nonzero_dir_acc"],
-    audit["quantitative_baselines_oos"]["2024"]["Momentum_1Bar"]["overall_nonzero_dir_acc"],
-    audit["quantitative_baselines_oos"]["2024"]["EMA_12_26_Crossover"]["overall_nonzero_dir_acc"],
-    audit["quantitative_baselines_oos"]["2024"]["MeanReversion_5Bar"]["overall_nonzero_dir_acc"],
-    audit["multi_checkpoint_oos_audit"]["epoch1"]["2024"]["overall_nonzero_dir_acc"],
-    audit["multi_checkpoint_oos_audit"]["epoch2"]["2024"]["overall_nonzero_dir_acc"],
-    audit["multi_checkpoint_oos_audit"]["epoch3"]["2024"]["overall_nonzero_dir_acc"]
-]
-xb = np.arange(len(models_lbl))
-ax2.bar(xb - 0.18, raw_vals, 0.35, label='Raw Directional Acc (%)', color='#1E3A8A', edgecolor='black', lw=0.5)
-ax2.bar(xb + 0.18, nz_vals, 0.35, label='Non-Zero Strict Acc (%)', color='#38BDF8', edgecolor='black', lw=0.5)
-ax2.axhline(50.0, color='#E11D48', linestyle='--', lw=1.2)
+# Panel (b): 1-Minute Accuracy & ROC-AUC by |y| Return Magnitude Quartile (Q1..Q4, y != 0)
+b_keys = ["Q1_Smallest_Move", "Q2_Moderate_Move", "Q3_Medium_Move", "Q4_Largest_Move"]
+b_lbls = ["Q1 Micro\n(~0.75 bps)", "Q2 Small\n(~2.1 bps)", "Q3 Medium\n(~4.4 bps)", "Q4 Large\n(~12.4 bps)"]
+cal_acc_24 = [ep3_24["1m"]["return_magnitude_buckets"][bk]["calibrated_dir_acc"] for bk in b_keys]
+cal_acc_26 = [ep3_26["1m"]["return_magnitude_buckets"][bk]["calibrated_dir_acc"] for bk in b_keys]
+b_auc_24 = [ep3_24["1m"]["return_magnitude_buckets"][bk]["roc_auc"] * 100.0 for bk in b_keys]
+b_auc_26 = [ep3_26["1m"]["return_magnitude_buckets"][bk]["roc_auc"] * 100.0 for bk in b_keys]
+
+xb = np.arange(len(b_lbls))
+wb = 0.20
+ax2.bar(xb - 1.5*wb, cal_acc_24, wb, label='2024 Calibrated Acc (%)', color='#1E3A8A', edgecolor='black', lw=0.5)
+ax2.bar(xb - 0.5*wb, b_auc_24, wb, label='2024 Bucket ROC-AUC (×100)', color='#38BDF8', edgecolor='black', lw=0.5)
+ax2.bar(xb + 0.5*wb, cal_acc_26, wb, label='2026 Calibrated Acc (%)', color='#0F766E', edgecolor='black', lw=0.5)
+ax2.bar(xb + 1.5*wb, b_auc_26, wb, label='2026 Bucket ROC-AUC (×100)', color='#F59E0B', edgecolor='black', lw=0.5)
+ax2.axhline(50.0, color='#E11D48', linestyle='--', lw=1.2, label='50.0% Null Benchmark')
 ax2.set_xticks(xb)
-ax2.set_xticklabels(models_lbl, fontsize=7.5)
-ax2.set_ylim(48.0, 53.2)
-ax2.set_ylabel('7-Horizon Mean Accuracy (%)')
-ax2.set_title('(b) Epoch Progression vs. Classical Baselines (2024 OOS)')
-ax2.legend(loc='upper left', frameon=True)
+ax2.set_xticklabels(b_lbls, fontsize=8)
+ax2.set_ylim(49.0, 54.8)
+ax2.set_ylabel('Non-Zero Bucket Accuracy (%) / AUC (×100)')
+ax2.set_title('(b) 1-Minute Non-Zero Discrimination by |Return| Quartile')
+ax2.legend(loc='upper right', frameon=True, fontsize=7.2)
 ax2.grid(axis='y', alpha=0.3)
 
 plt.tight_layout()
@@ -141,146 +133,159 @@ fig.savefig("papers/figures/fig2_oos_directional_accuracy.png", dpi=300, bbox_in
 plt.close()
 
 # =========================================================================
-# Figure 3: Exact Frobenius Weight Drift & Walk-Forward ROC-AUC / IC Transfer
+# Figure 3: Frobenius Drift vs RMS Update Arithmetic & Pre-2024 Stress Forgetting
 # =========================================================================
-print("Rendering Figure 3 from real Frobenius weight drift & walk-forward telemetry...")
+print("Rendering Figure 3 from Frobenius drift / Adam RMS arithmetic & Pre-2024 stress audit...")
 fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(11, 4.0), dpi=300)
 
-modules = ["Return\nHead", "Conformer\nBackbone", "Directional\nHead", "Inception\nTokenizer"]
-mod_keys = ["return_regression_head", "norms_and_aux", "directional_class_head", "inception_tokenizer"]
+modules = ["Return\nHead", "Depthwise\nConv", "SwiGLU\nFFN", "Conformer\nBackbone", "Dir.\nHead", "GQA\nAttn", "Inception\nTokenizer"]
+mod_keys = ["return_regression_head", "sub_depthwise_conv", "sub_swiglu_ffn", "conformer_backbone_all", "directional_class_head", "sub_gqa_attention", "inception_tokenizer"]
 
-drift_ep1_ep2 = [audit["weight_drift_frobenius_audit"]["epoch1_to_epoch2"]["by_module"][k]["rel_frobenius_pct"] for k in mod_keys]
-drift_ep2_ep3 = [audit["weight_drift_frobenius_audit"]["epoch2_to_epoch3"]["by_module"][k]["rel_frobenius_pct"] for k in mod_keys]
-drift_wf24 = [audit["weight_drift_frobenius_audit"]["epoch3_to_wf2024"]["by_module"][k]["rel_frobenius_pct"] for k in mod_keys]
-drift_wf25 = [audit["weight_drift_frobenius_audit"]["epoch3_to_wf2025"]["by_module"][k]["rel_frobenius_pct"] for k in mod_keys]
+wd = audit["weight_drift_and_cosine_audit"]
+drift_wf24 = [wd["epoch3_to_wf2024"]["by_module"][k]["rel_frobenius_pct"] for k in mod_keys]
+drift_wf25 = [wd["epoch3_to_wf2025"]["by_module"][k]["rel_frobenius_pct"] for k in mod_keys]
+rms_dw_wf24 = [wd["epoch3_to_wf2024"]["by_module"][k]["rms_update_magnitude"] * 1e4 for k in mod_keys]
+rms_w_base = [wd["epoch3_to_wf2024"]["by_module"][k]["rms_weight_norm"] * 50.0 for k in mod_keys]
 
 xm = np.arange(len(modules))
-wm = 0.2
-ax1.bar(xm - 1.5*wm, drift_ep1_ep2, wm, label='Pre-Train Ep1→Ep2 (10.00% Global)', color='#64748B', edgecolor='black', lw=0.5)
-ax1.bar(xm - 0.5*wm, drift_ep2_ep3, wm, label='Pre-Train Ep2→Ep3 (5.73% Global)', color='#94A3B8', edgecolor='black', lw=0.5)
-ax1.bar(xm + 0.5*wm, drift_wf24, wm, label='Walk-Forward Ep3→WF24 (0.54% Global)', color='#0EA5E9', edgecolor='black', lw=0.5)
-ax1.bar(xm + 1.5*wm, drift_wf25, wm, label='Walk-Forward Ep3→WF25 (0.79% Global)', color='#0F766E', edgecolor='black', lw=0.5)
-ax1.set_yscale('log')
+wm = 0.25
+ax1.bar(xm - wm, drift_wf24, wm, label='Rel. Drift Ep3→WF24 (%)', color='#0284C7', edgecolor='black', lw=0.5)
+ax1.bar(xm, drift_wf25, wm, label='Rel. Drift Ep3→WF25 (%)', color='#0D9488', edgecolor='black', lw=0.5)
+ax1.bar(xm + wm, rms_dw_wf24, wm, label='Abs. RMS Update ||ΔW||_RMS (×10⁴)', color='#F59E0B', edgecolor='black', lw=0.5)
+ax1.plot(xm, rms_w_base, marker='D', color='#E11D48', lw=1.5, linestyle='--', label='Base Weight Norm ||W||_RMS (×50)')
 ax1.set_xticks(xm)
-ax1.set_xticklabels(modules)
-ax1.set_ylabel('Relative Frobenius Drift ||ΔW||_F / ||W||_F (%)')
-ax1.set_title('(a) Layer-Wise Parameter Drift Hierarchy (Log Scale)')
-ax1.legend(loc='upper left', frameon=True, fontsize=7.5)
-ax1.grid(True, which="both", alpha=0.3)
-
-# Panel (b): 2025 OOS ROC-AUC across Checkpoints (Epoch 1, Epoch 2, Epoch 3, WF-2024, WF-2025)
-ckpts = ["Epoch 1\n(Zero-Shot)", "Epoch 2\n(Zero-Shot)", "Epoch 3\n(Zero-Shot)", "WF-2024\n(Online)", "WF-2025\n(Online)"]
-ck_keys = ["epoch1", "epoch2", "epoch3", "wf2024", "wf2025"]
-aucs_2025 = [audit["multi_checkpoint_oos_audit"][k]["2025"]["overall_roc_auc"] for k in ck_keys]
-aucs_2024 = [audit["multi_checkpoint_oos_audit"][k]["2024"]["overall_roc_auc"] for k in ck_keys]
-
-xc = np.arange(len(ckpts))
-ax2.plot(xc, aucs_2024, marker='o', lw=2.0, color='#1D4ED8', label='2024 OOS Mean ROC-AUC')
-ax2.plot(xc, aucs_2025, marker='s', lw=2.0, color='#0D9488', linestyle='--', label='2025 OOS Mean ROC-AUC')
-ax2.axhline(0.5000, color='#E11D48', linestyle=':', lw=1.2, label='Random Classifier (0.5000)')
+ax1.set_xticklabels(modules, fontsize=7.5)
+ax1.set_ylabel('Drift (%) / Scaled RMS Magnitude')
+ax1.set_title('(a) Plasticity Hierarchy: Relative Drift vs. Absolute RMS Update')
+ax1.legend(loc='upper left', frameon=True, fontsize=7.2)
+ax1.grid(True, alpha=0.3)
+
+# Panel (b): Pre-2024 Historical Stress Regimes & Forward OOS ROC-AUC (Epoch 3 vs WF-2024 vs WF-2025)
+cohorts_lbl = ["2020-03\nCOVID Crash", "2022-06\nRate Shock", "2023-12\nPre-Anchor", "Pre-2024\nCombined", "2024\nOOS", "2025\nOOS", "2026\nForward"]
+c_keys = ["stress_2020_03", "stress_2022_06", "stress_2023_12", "pre2024_stress", "2024", "2025", "2026"]
+
+mc = audit["multi_checkpoint_oos_and_stress_audit"]
+auc_ep3 = [mc["epoch3"][ck]["overall_roc_auc"] for ck in c_keys]
+auc_wf24 = [mc["wf2024"][ck]["overall_roc_auc"] for ck in c_keys]
+auc_wf25 = [mc["wf2025"][ck]["overall_roc_auc"] for ck in c_keys]
+
+xc = np.arange(len(cohorts_lbl))
+wc = 0.25
+ax2.bar(xc - wc, auc_ep3, wc, label='Frozen Epoch 3 (2010–2023)', color='#1E3A8A', edgecolor='black', lw=0.5)
+ax2.bar(xc, auc_wf24, wc, label='Online WF-2024 (+2024)', color='#0EA5E9', edgecolor='black', lw=0.5)
+ax2.bar(xc + wc, auc_wf25, wc, label='Online WF-2025 (+2024–25)', color='#10B981', edgecolor='black', lw=0.5)
+ax2.axhline(0.5000, color='#E11D48', linestyle='--', lw=1.1, label='Random Null (0.5000)')
 ax2.set_xticks(xc)
-ax2.set_xticklabels(ckpts, fontsize=8)
-ax2.set_ylim(0.508, 0.520)
+ax2.set_xticklabels(cohorts_lbl, fontsize=7.5)
+ax2.set_ylim(0.500, 0.522)
 ax2.set_ylabel('7-Horizon Mean ROC-AUC')
-ax2.set_title('(b) Out-of-Sample ROC-AUC Across Checkpoints')
-ax2.legend(loc='lower right', frameon=True)
-ax2.grid(True, alpha=0.3)
+ax2.set_title('(b) Backward Stress Forgetting & Forward Transfer Audit')
+ax2.legend(loc='upper right', frameon=True, fontsize=7.2)
+ax2.grid(axis='y', alpha=0.3)
 
 plt.tight_layout()
 fig.savefig("papers/figures/fig3_walk_forward_comparison.png", dpi=300, bbox_inches='tight', pad_inches=0.1)
 plt.close()
 
 # =========================================================================
-# Figure 4: Mechanistic Inception Kernel Ablation (Continuous L2 & Accuracy)
+# Figure 4: Controlled Mechanistic Kernel Ablation (Mean-Ablation, Controls & Denominator Effect)
 # =========================================================================
-print("Rendering Figure 4 from real 7-horizon mechanistic kernel ablation...")
+print("Rendering Figure 4 from controlled mean-ablation, random-channel controls & RMS logit shift...")
 fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(11, 4.0), dpi=300)
 
-ab_branches = ["Mask_K3", "Mask_K7", "Mask_K15", "Mask_K31", "Mask_K63", "Mask_K127", "Mask_Micro_K3_K7", "Mask_Meso_K15_K31", "Mask_Macro_K63_K127"]
-ab_labels = ["K=3", "K=7", "K=15", "K=31", "K=63", "K=127", "Micro\n(3,7)", "Meso\n(15,31)", "Macro\n(63,127)"]
+ab = audit["mechanistic_kernel_ablation_controlled"]
+ab_keys = ["MeanAblate_K3", "MeanAblate_K7", "MeanAblate_K15", "MeanAblate_K31", "MeanAblate_K63", "MeanAblate_K127", "Random_341ch_Control", "MeanAblate_Micro_K3_K7", "MeanAblate_Meso_K15_K31", "MeanAblate_Macro_K63_K127", "Random_682ch_Control"]
+ab_lbls = ["K=3", "K=7", "K=15", "K=31", "K=63", "K=127", "Rand\n341ch", "Micro\n(3,7)", "Meso\n(15,31)", "Macro\n(63,127)", "Rand\n682ch"]
 
-logit_l2_1m = [audit["mechanistic_kernel_ablation_7h"]["ablations"][b]["per_horizon"]["1m"]["logit_l2_perturbation_pct"] for b in ab_branches]
-logit_l2_15m = [audit["mechanistic_kernel_ablation_7h"]["ablations"][b]["per_horizon"]["15m"]["logit_l2_perturbation_pct"] for b in ab_branches]
-logit_l2_1d = [audit["mechanistic_kernel_ablation_7h"]["ablations"][b]["per_horizon"]["1d"]["logit_l2_perturbation_pct"] for b in ab_branches]
+cent_mean = [ab["mean_ablations"][k]["mean_centered_logit_l2_pct"] for k in ab_keys]
+cent_zero = [ab["zero_ablations"][k.replace("MeanAblate_", "ZeroAblate_")]["mean_centered_logit_l2_pct"] for k in ab_keys]
 
-xa = np.arange(len(ab_labels))
-wa = 0.26
-ax1.bar(xa - wa, logit_l2_1m, wa, label='1m Horizon Logit L2 Shift (%)', color='#0284C7', edgecolor='black', lw=0.5)
-ax1.bar(xa, logit_l2_15m, wa, label='15m Horizon Logit L2 Shift (%)', color='#F59E0B', edgecolor='black', lw=0.5)
-ax1.bar(xa + wa, logit_l2_1d, wa, label='1d Horizon Logit L2 Shift (%)', color='#10B981', edgecolor='black', lw=0.5)
+xa = np.arange(len(ab_lbls))
+wa = 0.36
+ax1.bar(xa - wa/2, cent_mean, wa, label='In-Distribution Mean-Ablation (Centered L2 %)', color='#4F46E5', edgecolor='black', lw=0.5)
+ax1.bar(xa + wa/2, cent_zero, wa, label='Zero-Clamping Ablation (Centered L2 %)', color='#0EA5E9', edgecolor='black', lw=0.5)
+ax1.axhline(ab["mean_ablations"]["Random_341ch_Control"]["mean_centered_logit_l2_pct"], color='#E11D48', linestyle=':', lw=1.2, label='Random 341-Ch Mean-Ablate Null (20.31%)')
 ax1.set_xticks(xa)
-ax1.set_xticklabels(ab_labels, fontsize=7.5)
-ax1.set_ylabel('Relative Logit Perturbation ||Δz||_2 / ||z||_2 (%)')
-ax1.set_title('(a) Continuous Logit Sensitivity by Receptive Field')
-ax1.legend(loc='upper left', frameon=True)
+ax1.set_xticklabels(ab_lbls, fontsize=7.2)
+ax1.set_ylabel('Centered Logit Shift ||Δ(z - z̄)||_2 / ||z - z̄||_2 (%)')
+ax1.set_title('(a) Centered Logit Sensitivity vs. Random-Channel Controls')
+ax1.legend(loc='upper left', frameon=True, fontsize=7.2)
 ax1.grid(axis='y', alpha=0.3)
 
-# Panel (b): Return Prediction L2 Shift (%) vs Mean Logit L2 Shift (%)
-mean_l2_z = [audit["mechanistic_kernel_ablation_7h"]["ablations"][b]["mean_logit_l2_perturbation_pct"] for b in ab_branches]
-mean_l2_r = [audit["mechanistic_kernel_ablation_7h"]["ablations"][b]["mean_return_l2_perturbation_pct"] for b in ab_branches]
-
-ax2.bar(xa - 0.18, mean_l2_z, 0.35, label='Mean Logit L2 Shift (%)', color='#4F46E5', edgecolor='black', lw=0.5)
-ax2.bar(xa + 0.18, mean_l2_r, 0.35, label='Mean Return Forecast L2 Shift (%)', color='#E11D48', edgecolor='black', lw=0.5)
-ax2.set_xticks(xa)
-ax2.set_xticklabels(ab_labels, fontsize=7.5)
-ax2.set_ylabel('7-Horizon Mean Relative L2 Perturbation (%)')
-ax2.set_title('(b) Dual-Head Perturbation Magnitude Under Masking')
-ax2.legend(loc='upper left', frameon=True)
-ax2.grid(axis='y', alpha=0.3)
+# Panel (b): Resolving the Denominator Effect across Horizons (Uncentered Raw L2 vs Centered L2 vs Abs RMS)
+k7_ph = ab["mean_ablations"]["MeanAblate_K7"]["per_horizon"]
+meso_ph = ab["mean_ablations"]["MeanAblate_Meso_K15_K31"]["per_horizon"]
+raw_l2_k7 = [k7_ph[h]["raw_logit_l2_pct"] for h in horizons]
+cent_l2_k7 = [k7_ph[h]["centered_logit_l2_pct"] for h in horizons]
+rms_k7_scaled = [k7_ph[h]["abs_logit_rms_shift"] * 1000.0 for h in horizons]
+base_rms_scaled = [ab["baseline"]["per_horizon"][h]["logit_rms"] * 200.0 for h in horizons]
+
+xh = np.arange(len(horizons))
+ax2.plot(xh, raw_l2_k7, marker='o', lw=2.0, color='#D97706', label='Uncentered ||Δz||_2 / ||z||_2 (%) [Peaks at 15m]')
+ax2.plot(xh, cent_l2_k7, marker='s', lw=2.0, color='#4F46E5', label='Centered ||Δ(z-z̄)||_2 / ||z-z̄||_2 (%) [Variance-Norm]')
+ax2.plot(xh, rms_k7_scaled, marker='^', lw=2.0, color='#0F766E', label='Denominator-Free RMS(Δz) (×1000) [Monotonic]')
+ax2.plot(xh, base_rms_scaled, marker='x', lw=1.4, linestyle='--', color='#64748B', label='Baseline Denom. ||z||_RMS (×200) [Min at 15m]')
+ax2.set_xticks(xh)
+ax2.set_xticklabels(horizons)
+ax2.set_ylabel('Perturbation (%) / Scaled RMS Units')
+ax2.set_title('(b) Decomposition of the 15m Logit Denominator Effect (K=7)')
+ax2.legend(loc='upper right', frameon=True, fontsize=7.0)
+ax2.grid(True, alpha=0.3)
 
 plt.tight_layout()
 fig.savefig("papers/figures/fig4_kernel_ablation.png", dpi=300, bbox_inches='tight', pad_inches=0.1)
 plt.close()
 
 # =========================================================================
-# Figure 5: Strictly Causal Chronological Execution Backtest & Hurdle Filtering
+# Figure 5: Non-Overlapping Contiguous Backtest & Frozen 2024 Policy OOS
 # =========================================================================
-print("Rendering Figure 5 from real causal chronological backtest...")
+print("Rendering Figure 5 from non-overlapping state-turnover backtest & frozen 2024 policy...")
 fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(11, 4.0), dpi=300)
 
-taus_lbl = ["0.0σ\n(100% Trd)", "0.5σ\n(~55% Trd)", "1.0σ\n(~20% Trd)", "1.5σ\n(~5% Trd)", "2.0σ\n(~1% Trd)"]
-tau_keys = ["tau_0.0sigma", "tau_0.5sigma", "tau_1.0sigma", "tau_1.5sigma", "tau_2.0sigma"]
-
-# Panel (a): 2025 60m & 2026 15m Breakeven Fee Capacitance (bps) vs Confidence Hurdle
-be_2025_60m = [audit["causal_chronological_backtest"]["2025"]["60m"][tk]["breakeven_fee_bps"] for tk in tau_keys]
-be_2026_15m = [audit["causal_chronological_backtest"]["2026"]["15m"][tk]["breakeven_fee_bps"] for tk in tau_keys]
-be_2024_5m  = [audit["causal_chronological_backtest"]["2024"]["5m"][tk]["breakeven_fee_bps"] for tk in tau_keys]
-
-xt = np.arange(len(taus_lbl))
-ax1.plot(xt, be_2025_60m, marker='o', lw=2.0, color='#0D9488', label='2025 60m Horizon Breakeven Fee')
-ax1.plot(xt, be_2026_15m, marker='s', lw=2.0, color='#2563EB', label='2026 15m Horizon Breakeven Fee')
-ax1.plot(xt, be_2024_5m, marker='^', lw=2.0, color='#D97706', label='2024 5m Horizon Breakeven Fee')
-ax1.axhline(1.0, color='#E11D48', linestyle='--', lw=1.2, label='Institutional Prime Fee (1.0 bps)')
-ax1.set_xticks(xt)
-ax1.set_xticklabels(taus_lbl, fontsize=8)
-ax1.set_ylabel('Breakeven One-Way Fee Capacitance (bps)')
-ax1.set_title('(a) Causal Confidence Hurdle (τ) vs. Breakeven Fee')
-ax1.legend(loc='upper left', frameon=True)
-ax1.grid(True, alpha=0.3)
+# Panel (a): Fixed Hurdle (tau=1.0sigma) Walk-Forward Gross Return & State-Hold Breakeven Across 2024, 2025, 2026
+wf_tau1 = audit["frozen_2024_policy_oos_evaluation"]["walk_forward_weights"]["fixed_tau_1sigma_all_horizons"]
+h_plot = ["1m", "3m", "5m", "15m", "30m", "60m"]
+be_24 = [wf_tau1[h]["2024"]["breakeven_one_way_bps_state_hold"] for h in h_plot]
+be_25 = [wf_tau1[h]["2025"]["breakeven_one_way_bps_state_hold"] for h in h_plot]
+be_26 = [wf_tau1[h]["2026"]["breakeven_one_way_bps_state_hold"] for h in h_plot]
+
+xp = np.arange(len(h_plot))
+wp = 0.25
+ax1.bar(xp - wp, be_24, wp, label='2024 OOS State-Hold Breakeven (bps)', color='#0284C7', edgecolor='black', lw=0.5)
+ax1.bar(xp, be_25, wp, label='2025 OOS State-Hold Breakeven (bps)', color='#0D9488', edgecolor='black', lw=0.5)
+ax1.bar(xp + wp, be_26, wp, label='2026 OOS State-Hold Breakeven (bps)', color='#7C3AED', edgecolor='black', lw=0.5)
+ax1.axhline(0.0, color='black', lw=0.9)
+ax1.axhline(1.5, color='#E11D48', linestyle='--', lw=1.2, label='Maker / Midpoint Cost Hurdle (1.5 bps)')
+ax1.set_xticks(xp)
+ax1.set_xticklabels(h_plot)
+ax1.set_ylabel('State-Hold Breakeven One-Way Cost (bps)')
+ax1.set_title('(a) Non-Overlapping Breakeven Cost (Fixed τ=1.0σ, Walk-Forward)')
+ax1.legend(loc='lower left', frameon=True, fontsize=7.2)
+ax1.grid(axis='y', alpha=0.3)
 
-# Panel (b): Net Post-Friction Sharpe Ratio across Slippage Grid (0, 1, 2, 5 bps)
-bps_grid = [0.0, 1.0, 2.0, 5.0]
-bps_keys = ["0.0bps", "1.0bps", "2.0bps", "5.0bps"]
-
-sh_2025_1d_t0 = [audit["causal_chronological_backtest"]["2025"]["1d"]["tau_0.0sigma"]["slippage_grid"][bk]["sharpe"] for bk in bps_keys]
-sh_2025_60m_t15 = [audit["causal_chronological_backtest"]["2025"]["60m"]["tau_1.5sigma"]["slippage_grid"][bk]["sharpe"] for bk in bps_keys]
-sh_2024_5m_t20 = [audit["causal_chronological_backtest"]["2024"]["5m"]["tau_2.0sigma"]["slippage_grid"][bk]["sharpe"] for bk in bps_keys]
-sh_2026_15m_t20 = [audit["causal_chronological_backtest"]["2026"]["15m"]["tau_2.0sigma"]["slippage_grid"][bk]["sharpe"] for bk in bps_keys]
-
-ax2.plot(bps_grid, sh_2024_5m_t20, marker='^', lw=2.0, color='#D97706', label='2024 5m (τ=2.0σ, 99.3% Turn. Saved)')
-ax2.plot(bps_grid, sh_2025_1d_t0, marker='o', lw=2.0, color='#1D4ED8', label='2025 1d (Unfiltered τ=0.0σ)')
-ax2.plot(bps_grid, sh_2025_60m_t15, marker='s', lw=2.0, color='#0D9488', label='2025 60m (τ=1.5σ, 94.4% Turn. Saved)')
-ax2.plot(bps_grid, sh_2026_15m_t20, marker='D', lw=2.0, color='#7C3AED', label='2026 15m (τ=2.0σ, 98.9% Turn. Saved)')
-ax2.axhline(0.0, color='black', linestyle='--', lw=1.1, label='Breakeven Sharpe (0.0)')
-ax2.set_xlabel('One-Way Execution Slippage / Fee (Basis Points)')
-ax2.set_ylabel('Annualized Net Sharpe Ratio (Causal Chronological)')
-ax2.set_title('(b) Post-Friction Net Sharpe Survival Under Hurdle')
-ax2.legend(loc='lower left', frameon=True, fontsize=7.5)
+# Panel (b): Frozen 2024 Selected Policy (60m, tau=0.0sigma) vs Fixed 3m (tau=1.0sigma) Across Cost Grid
+cost_bps = [0.0, 1.5, 3.0, 5.0]
+cost_keys = ["0.0bps_oneway", "1.5bps_oneway", "3.0bps_oneway", "5.0bps_oneway"]
+
+fpol_wf = audit["frozen_2024_policy_oos_evaluation"]["walk_forward_weights"]["frozen_2024_optimal_policy"]["yearly_performance"]
+sr_f24 = [fpol_wf["2024"]["cost_grid"][ck]["annualized_sharpe_state_hold"] for ck in cost_keys]
+sr_f25 = [fpol_wf["2025"]["cost_grid"][ck]["annualized_sharpe_state_hold"] for ck in cost_keys]
+sr_f26 = [fpol_wf["2026"]["cost_grid"][ck]["annualized_sharpe_state_hold"] for ck in cost_keys]
+sr_3m_26 = [wf_tau1["3m"]["2026"]["cost_grid"][ck]["annualized_sharpe_state_hold"] for ck in cost_keys]
+
+ax2.plot(cost_bps, sr_f24, marker='o', lw=2.0, color='#0284C7', label='Frozen Policy on 2024 (Selection: SR=+4.40 at 0bp, +2.04 at 1.5bp)')
+ax2.plot(cost_bps, sr_f25, marker='s', lw=2.0, color='#0D9488', label='Frozen Policy Applied Untouched to 2025 (SR=+1.01 at 0bp)')
+ax2.plot(cost_bps, sr_f26, marker='^', lw=2.0, color='#D97706', label='Frozen Policy Applied Untouched to 2026 (SR=-2.01 at 0bp)')
+ax2.plot(cost_bps, sr_3m_26, marker='D', lw=1.8, linestyle='--', color='#7C3AED', label='2026 3m Demeaned (τ=1.0σ: SR=+3.56 at 0bp)')
+ax2.axhline(0.0, color='black', linestyle='--', lw=1.1, label='Zero Net Sharpe Boundary')
+ax2.set_xlabel('One-Way Execution Cost: Spread + Fee + Slippage (bps)')
+ax2.set_ylabel('Annualized Net Sharpe Ratio (Non-Overlapping)')
+ax2.set_title('(b) Frozen 2024 Policy Untouched OOS Transfer (2024–2026)')
+ax2.legend(loc='lower left', frameon=True, fontsize=7.0)
 ax2.grid(True, alpha=0.3)
 
 plt.tight_layout()
 fig.savefig("papers/figures/fig5_slippage_friction_curve.png", dpi=300, bbox_inches='tight', pad_inches=0.1)
 plt.close()
 
-print("✅ All 5 deterministic empirical figures rendered at 300 DPI!")
-
+print("✅ All 5 upgraded deterministic empirical figures rendered at 300 DPI!")