agent-os-kernel 1.1.0__py3-none-any.whl → 1.2.0__py3-none-any.whl

modules/scak/paper/figures/generate_figures.py

@@ -0,0 +1,463 @@
+"""
+Generate all 6 figures for the SCAK paper.
+
+Figures:
+1. fig1_ooda_architecture - Dual-Loop OODA Architecture
+2. fig2_memory_hierarchy - Three-Tier Memory Hierarchy
+3. fig3_gaia_results - GAIA Benchmark Bar Chart
+4. fig4_ablation_heatmap - Ablation Study Heatmap
+5. fig5_context_reduction - Context Token Reduction Line Chart
+6. fig6_mttr_boxplot - Chaos Engineering MTTR Box Plot
+"""
+
+import matplotlib.pyplot as plt
+import matplotlib.patches as patches
+import numpy as np
+import os
+
+# Output directory (relative to script location)
+SCRIPT_DIR = os.path.dirname(os.path.abspath(__file__))
+OUTPUT_DIR = SCRIPT_DIR
+
+# Set global style for academic figures
+plt.rcParams['font.family'] = 'sans-serif'
+plt.rcParams['font.sans-serif'] = ['Arial', 'DejaVu Sans']
+plt.rcParams['font.size'] = 10
+plt.rcParams['axes.titlesize'] = 12
+plt.rcParams['axes.labelsize'] = 10
+
+
+def draw_box(ax, xy, width, height, label, color='#e0e0e0', edge='black', fontsize=9):
+    """Helper to draw a styled box with label."""
+    box = patches.FancyBboxPatch(
+        xy, width, height,
+        boxstyle='round,pad=0.05',
+        linewidth=1.5,
+        edgecolor=edge,
+        facecolor=color,
+        zorder=2
+    )
+    ax.add_patch(box)
+    ax.text(
+        xy[0] + width/2, xy[1] + height/2, label,
+        ha='center', va='center', fontweight='bold',
+        fontsize=fontsize, zorder=3, color='black'
+    )
+    return box
+
+
+def create_fig1_ooda_architecture():
+    """Figure 1: Dual-Loop OODA Architecture."""
+    fig, ax = plt.subplots(figsize=(12, 8))
+    ax.set_xlim(0, 12)
+    ax.set_ylim(0, 10)
+    ax.axis('off')
+    
+    # User Query at top
+    draw_box(ax, (4.5, 9.0), 3, 0.7, "USER QUERY", color='#ffffff', edge='#333')
+    
+    # Loop 1: Runtime (Blue background)
+    loop1_bg = patches.FancyBboxPatch(
+        (0.5, 5.0), 11, 3.5,
+        boxstyle='round,pad=0.1',
+        linewidth=2,
+        edgecolor='#1565C0',
+        facecolor='#E3F2FD',
+        zorder=0
+    )
+    ax.add_patch(loop1_bg)
+    ax.text(1.0, 8.2, "LOOP 1: RUNTIME (OBSERVE → ACT)", fontsize=11, 
+            fontweight='bold', color='#0D47A1')
+    
+    # Loop 1 components
+    draw_box(ax, (1.0, 6.5), 2.2, 1.2, "TRIAGE\nENGINE\n(Sync/Async)", 
+             color='#FF9800', edge='#E65100', fontsize=8)
+    draw_box(ax, (4.0, 6.5), 2.2, 1.2, "EXECUTE\nAGENT\n(GPT-4o)", 
+             color='#FFCC80', edge='#E65100', fontsize=8)
+    draw_box(ax, (7.0, 6.5), 2.2, 1.2, "RESPOND\n\n(Return)", 
+             color='#C8E6C9', edge='#2E7D32', fontsize=8)
+    draw_box(ax, (4.0, 5.2), 2.2, 1.0, "GIVE-UP\nDETECTOR", 
+             color='#FFCDD2', edge='#C62828', fontsize=8)
+    
+    # Loop 2: Alignment (Green background)
+    loop2_bg = patches.FancyBboxPatch(
+        (0.5, 0.5), 11, 4.0,
+        boxstyle='round,pad=0.1',
+        linewidth=2,
+        edgecolor='#2E7D32',
+        facecolor='#E8F5E9',
+        zorder=0
+    )
+    ax.add_patch(loop2_bg)
+    ax.text(1.0, 4.2, "LOOP 2: ALIGNMENT (ORIENT → DECIDE)", fontsize=11, 
+            fontweight='bold', color='#1B5E20')
+    
+    # Loop 2 components
+    draw_box(ax, (1.0, 2.5), 2.2, 1.2, "COMPLETENESS\nAUDITOR\n(Diff Audit)", 
+             color='#E1BEE7', edge='#7B1FA2', fontsize=8)
+    draw_box(ax, (4.0, 2.5), 2.2, 1.2, "SHADOW\nTEACHER\n(o1-preview)", 
+             color='#9C27B0', edge='#4A148C', fontsize=8)
+    draw_box(ax, (7.0, 2.5), 2.2, 1.2, "MEMORY\nCONTROLLER\n(Tiered)", 
+             color='#009688', edge='#004D40', fontsize=8)
+    draw_box(ax, (4.0, 0.8), 2.2, 1.2, "GAP ANALYSIS\n→ PATCH", 
+             color='#B2DFDB', edge='#00695C', fontsize=8)
+    draw_box(ax, (7.0, 0.8), 2.2, 1.0, "APPLY PATCH\nto Memory", 
+             color='#80CBC4', edge='#00695C', fontsize=8)
+    
+    # Arrows
+    arrow_props = dict(arrowstyle='->', lw=2, color='#424242')
+    
+    # User -> Triage
+    ax.annotate("", xy=(2.1, 8.5), xytext=(5.0, 9.0), arrowprops=arrow_props)
+    # Triage -> Execute
+    ax.annotate("", xy=(4.0, 7.1), xytext=(3.2, 7.1), arrowprops=arrow_props)
+    # Execute -> Respond
+    ax.annotate("", xy=(7.0, 7.1), xytext=(6.2, 7.1), arrowprops=arrow_props)
+    # Respond -> User
+    ax.annotate("", xy=(10.5, 8.5), xytext=(8.5, 7.7), arrowprops=arrow_props)
+    ax.text(10.2, 8.0, "User", fontsize=9)
+    # Execute -> Give-up detector
+    ax.annotate("", xy=(5.1, 6.5), xytext=(5.1, 6.2), arrowprops=arrow_props)
+    # Give-up -> Auditor (async)
+    ax.annotate("", xy=(3.2, 3.7), xytext=(5.1, 5.2), 
+                arrowprops=dict(arrowstyle='->', lw=2, color='#F44336', linestyle='dashed'))
+    ax.text(3.5, 4.6, "5-10%\n(async)", fontsize=8, color='#F44336', ha='center')
+    
+    # Loop 2 internal arrows
+    ax.annotate("", xy=(4.0, 3.1), xytext=(3.2, 3.1), arrowprops=arrow_props)
+    ax.annotate("", xy=(7.0, 3.1), xytext=(6.2, 3.1), arrowprops=arrow_props)
+    ax.annotate("", xy=(5.1, 2.5), xytext=(5.1, 2.0), arrowprops=arrow_props)
+    ax.annotate("", xy=(7.0, 1.3), xytext=(6.2, 1.4), arrowprops=arrow_props)
+    
+    plt.title("Figure 1: SCAK Dual-Loop Architecture", fontweight='bold', fontsize=14, pad=20)
+    plt.tight_layout()
+    plt.savefig(os.path.join(OUTPUT_DIR, 'fig1_ooda_architecture.png'), dpi=300, bbox_inches='tight')
+    plt.savefig(os.path.join(OUTPUT_DIR, 'fig1_ooda_architecture.pdf'), bbox_inches='tight')
+    plt.close()
+    print("Created fig1_ooda_architecture.png/pdf")
+
+
+def create_fig2_memory_hierarchy():
+    """Figure 2: Three-Tier Memory Hierarchy with Type A/B Lifecycle."""
+    fig, ax = plt.subplots(figsize=(12, 8))
+    ax.set_xlim(0, 12)
+    ax.set_ylim(0, 10)
+    ax.axis('off')
+    
+    # Tier 1: Kernel (top, red/orange)
+    draw_box(ax, (4.5, 8.0), 3, 1.5, "TIER 1: KERNEL\n(System Prompt)\n\n500 tokens\nALWAYS active", 
+             color='#FFCDD2', edge='#C62828', fontsize=9)
+    
+    # Tier 2: Cache (left)
+    draw_box(ax, (1.0, 5.5), 3, 1.8, "TIER 2: CACHE\n(Redis)\n\n10K entries\nConditional\nTool-specific", 
+             color='#FFF9C4', edge='#F9A825', fontsize=9)
+    
+    # Promotion/Demotion (center)
+    draw_box(ax, (4.5, 5.5), 3, 1.8, "PROMOTION\n/DEMOTION\n\n>10 hits/week → ↑\n<1 hit/month → ↓", 
+             color='#E0E0E0', edge='#616161', fontsize=9)
+    
+    # Tier 3: Archive (right)
+    draw_box(ax, (8.0, 5.5), 3, 1.8, "TIER 3: ARCHIVE\n(Vector DB)\n\nUnlimited\nOn-demand RAG\nLong-tail", 
+             color='#BBDEFB', edge='#1565C0', fontsize=9)
+    
+    # Arrows from Tier 1
+    ax.annotate("", xy=(2.5, 7.3), xytext=(4.5, 8.5), arrowprops=dict(arrowstyle='->', lw=2, color='#666'))
+    ax.annotate("", xy=(6.0, 7.3), xytext=(6.0, 8.0), arrowprops=dict(arrowstyle='->', lw=2, color='#666'))
+    ax.annotate("", xy=(9.5, 7.3), xytext=(7.5, 8.5), arrowprops=dict(arrowstyle='->', lw=2, color='#666'))
+    
+    # Bidirectional arrows between tiers
+    ax.annotate("", xy=(4.5, 6.4), xytext=(4.0, 6.4), arrowprops=dict(arrowstyle='<->', lw=2, color='#666'))
+    ax.annotate("", xy=(8.0, 6.4), xytext=(7.5, 6.4), arrowprops=dict(arrowstyle='<->', lw=2, color='#666'))
+    
+    # Type A/B boxes at bottom
+    # Type A (left)
+    type_a_bg = patches.FancyBboxPatch(
+        (0.5, 0.5), 5, 4.0,
+        boxstyle='round,pad=0.1',
+        linewidth=2,
+        edgecolor='#9E9E9E',
+        facecolor='#ECEFF1',
+        zorder=0
+    )
+    ax.add_patch(type_a_bg)
+    ax.text(3.0, 4.2, "TYPE A PATCHES (Syntax/Capability)", fontsize=10, 
+            fontweight='bold', color='#616161', ha='center')
+    ax.text(3.0, 3.5, "• \"Output valid JSON\"\n• \"Use ISO 8601 dates\"\n• \"Limit to 10 results\"", 
+            fontsize=9, ha='center', va='top')
+    ax.text(3.0, 1.8, "HIGH DECAY\nDelete on model upgrade\n~50 patches purged", 
+            fontsize=9, ha='center', va='top', color='#D32F2F', fontweight='bold')
+    
+    # Type B (right)
+    type_b_bg = patches.FancyBboxPatch(
+        (6.5, 0.5), 5, 4.0,
+        boxstyle='round,pad=0.1',
+        linewidth=2,
+        edgecolor='#388E3C',
+        facecolor='#C8E6C9',
+        zorder=0
+    )
+    ax.add_patch(type_b_bg)
+    ax.text(9.0, 4.2, "TYPE B PATCHES (Business/Context)", fontsize=10, 
+            fontweight='bold', color='#1B5E20', ha='center')
+    ax.text(9.0, 3.5, "• \"Fiscal year: July 1\"\n• \"Project_Alpha: archived\"\n• \"VIP users: priority\"", 
+            fontsize=9, ha='center', va='top')
+    ax.text(9.0, 1.8, "ZERO DECAY\nRetain indefinitely\n~10 patches kept", 
+            fontsize=9, ha='center', va='top', color='#1B5E20', fontweight='bold')
+    
+    plt.title("Figure 2: Three-Tier Memory Hierarchy with Type A/B Lifecycle", 
+              fontweight='bold', fontsize=14, pad=20)
+    plt.tight_layout()
+    plt.savefig(os.path.join(OUTPUT_DIR, 'fig2_memory_hierarchy.png'), dpi=300, bbox_inches='tight')
+    plt.savefig(os.path.join(OUTPUT_DIR, 'fig2_memory_hierarchy.pdf'), bbox_inches='tight')
+    plt.close()
+    print("Created fig2_memory_hierarchy.png/pdf")
+
+
+def create_fig3_gaia_results():
+    """Figure 3: GAIA Benchmark Results Bar Chart."""
+    # Data from spec
+    methods = ['GPT-4o\n(baseline)', 'AutoGen', 'LangGraph', 'o1-preview\nalone', 'Self-\nCritique', 'SCAK\n(ours)']
+    correction_rates = [8, 15, 0, 40, 40, 72]
+    post_patch_success = [8, 18, 5, 45, 48, 82]
+    
+    colors = ['#BDBDBD', '#FF9800', '#2196F3', '#9C27B0', '#009688', '#4CAF50']
+    
+    fig, ax = plt.subplots(figsize=(10, 6))
+    
+    x = np.arange(len(methods))
+    width = 0.35
+    
+    # Bars
+    bars1 = ax.bar(x - width/2, correction_rates, width, label='Correction Rate', 
+                   color=colors, edgecolor='black', linewidth=1)
+    bars2 = ax.bar(x + width/2, post_patch_success, width, label='Post-Patch Success',
+                   color=[c + '80' for c in colors], edgecolor='black', linewidth=1, alpha=0.7)
+    
+    # Add SCAK highlight border
+    bars1[-1].set_edgecolor('#FFD700')
+    bars1[-1].set_linewidth(3)
+    bars2[-1].set_edgecolor('#FFD700')
+    bars2[-1].set_linewidth(3)
+    
+    # Error bar for SCAK only
+    ax.errorbar(x[-1] - width/2, correction_rates[-1], yerr=4.2, fmt='none', 
+                color='black', capsize=5, capthick=2)
+    ax.errorbar(x[-1] + width/2, post_patch_success[-1], yerr=3.1, fmt='none', 
+                color='black', capsize=5, capthick=2)
+    
+    # Labels and formatting
+    ax.set_ylabel('Rate (%)', fontsize=12)
+    ax.set_xlabel('Method', fontsize=12)
+    ax.set_title('GAIA Laziness Benchmark: Correction Rate Comparison', fontweight='bold', fontsize=14)
+    ax.set_xticks(x)
+    ax.set_xticklabels(methods)
+    ax.set_ylim(0, 100)
+    ax.legend(loc='upper left')
+    ax.grid(axis='y', linestyle='--', alpha=0.3)
+    
+    # Add significance stars for SCAK
+    ax.text(x[-1], 88, '***', ha='center', fontsize=14, fontweight='bold')
+    ax.text(x[-1], 92, 'p<0.001', ha='center', fontsize=8)
+    
+    # Value labels on bars
+    for i, bar1 in enumerate(bars1):
+        ax.text(bar1.get_x() + bar1.get_width()/2, bar1.get_height() + 2, 
+                f'{int(correction_rates[i])}%', ha='center', fontsize=8, fontweight='bold')
+    
+    plt.tight_layout()
+    plt.savefig(os.path.join(OUTPUT_DIR, 'fig3_gaia_results.png'), dpi=300, bbox_inches='tight')
+    plt.savefig(os.path.join(OUTPUT_DIR, 'fig3_gaia_results.pdf'), bbox_inches='tight')
+    plt.close()
+    print("Created fig3_gaia_results.png/pdf")
+
+
+def create_fig4_ablation_heatmap():
+    """Figure 4: Ablation Study Heatmap."""
+    import matplotlib.colors as mcolors
+    
+    # Data from spec
+    configs = ['Full SCAK', '− Semantic Purge', '− Teacher (o1)', 
+               '− Tiered Memory', '− Diff Audit', 'Self-Critique Only']
+    detection = [100, 100, 45, 92, 0, 100]
+    correction = [72, 68, 28, 55, 0, 40]
+    
+    fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(12, 6))
+    
+    # Custom colormap: green (high) -> yellow (mid) -> red (low)
+    cmap = mcolors.LinearSegmentedColormap.from_list('custom', 
+           ['#E57373', '#FFB74D', '#FFF176', '#81C784', '#1B5E20'])
+    
+    # Detection Rate bars
+    bars1 = ax1.barh(configs, detection, color=[cmap(d/100) for d in detection], edgecolor='black')
+    ax1.set_xlabel('Detection Rate (%)', fontsize=12)
+    ax1.set_title('Detection Rate', fontweight='bold', fontsize=12)
+    ax1.set_xlim(0, 110)
+    ax1.axvline(x=100, color='green', linestyle='--', alpha=0.5)
+    
+    # Add value labels
+    for i, (bar, val) in enumerate(zip(bars1, detection)):
+        ax1.text(val + 2, bar.get_y() + bar.get_height()/2, f'{val}%', 
+                va='center', fontsize=10, fontweight='bold')
+    
+    # Correction Rate bars
+    bars2 = ax2.barh(configs, correction, color=[cmap(c/100) for c in correction], edgecolor='black')
+    ax2.set_xlabel('Correction Rate (%)', fontsize=12)
+    ax2.set_title('Correction Rate', fontweight='bold', fontsize=12)
+    ax2.set_xlim(0, 85)
+    ax2.axvline(x=72, color='green', linestyle='--', alpha=0.5)
+    
+    # Add value labels and significance
+    significance = ['', '*', '***', '**', '***', '***']
+    for i, (bar, val, sig) in enumerate(zip(bars2, correction, significance)):
+        label = f'{val}%'
+        if sig:
+            label += f' {sig}'
+        ax2.text(val + 2, bar.get_y() + bar.get_height()/2, label, 
+                va='center', fontsize=10, fontweight='bold')
+    
+    # Highlight Full SCAK row
+    for bar in [bars1[0], bars2[0]]:
+        bar.set_edgecolor('#FFD700')
+        bar.set_linewidth(3)
+    
+    plt.suptitle('Figure 4: Ablation Study - Component Contributions', 
+                 fontweight='bold', fontsize=14, y=1.02)
+    plt.tight_layout()
+    plt.savefig(os.path.join(OUTPUT_DIR, 'fig4_ablation_heatmap.png'), dpi=300, bbox_inches='tight')
+    plt.savefig(os.path.join(OUTPUT_DIR, 'fig4_ablation_heatmap.pdf'), bbox_inches='tight')
+    plt.close()
+    print("Created fig4_ablation_heatmap.png/pdf")
+
+
+def create_fig5_context_reduction():
+    """Figure 5: Context Token Reduction Line Chart."""
+    # Data from spec
+    time_points = ['T0', '+10', '+20', '+30', '+40', '+50', 'UPG', '+60', '+70']
+    no_purge = [800, 960, 1120, 1280, 1440, 1600, 1600, 1760, 1920]
+    with_scak = [800, 960, 1120, 1280, 1440, 1600, 880, 1040, 1200]
+    
+    fig, ax = plt.subplots(figsize=(10, 6))
+    
+    x = np.arange(len(time_points))
+    
+    # Lines
+    ax.plot(x, no_purge, 'o-', color='#F44336', linewidth=2, markersize=8, 
+            label='No Purge (unbounded)')
+    ax.plot(x, with_scak, 's-', color='#4CAF50', linewidth=2, markersize=8, 
+            label='SCAK (with Semantic Purge)')
+    
+    # Vertical line at model upgrade
+    upgrade_idx = 6
+    ax.axvline(x=upgrade_idx, color='#2196F3', linestyle='--', linewidth=2, alpha=0.7)
+    ax.text(upgrade_idx + 0.1, 1800, 'Model Upgrade\n(GPT-4o → GPT-5)', 
+            fontsize=10, color='#2196F3', fontweight='bold')
+    
+    # Shade the reduction area
+    ax.fill_between(x[upgrade_idx:], with_scak[upgrade_idx:], no_purge[upgrade_idx:], 
+                   alpha=0.2, color='#BBDEFB')
+    
+    # Annotation for 45% reduction
+    ax.annotate('45% reduction', xy=(7, 1050), xytext=(8, 1400),
+               arrowprops=dict(arrowstyle='->', color='#1565C0', lw=2),
+               fontsize=11, fontweight='bold', color='#1565C0')
+    
+    ax.set_xlabel('Patches Added', fontsize=12)
+    ax.set_ylabel('Context Tokens', fontsize=12)
+    ax.set_title('Context Token Growth: With and Without Semantic Purge', 
+                fontweight='bold', fontsize=14)
+    ax.set_xticks(x)
+    ax.set_xticklabels(time_points)
+    ax.set_ylim(600, 2100)
+    ax.legend(loc='upper left')
+    ax.grid(True, linestyle='--', alpha=0.3)
+    
+    plt.tight_layout()
+    plt.savefig(os.path.join(OUTPUT_DIR, 'fig5_context_reduction.png'), dpi=300, bbox_inches='tight')
+    plt.savefig(os.path.join(OUTPUT_DIR, 'fig5_context_reduction.pdf'), bbox_inches='tight')
+    plt.close()
+    print("Created fig5_context_reduction.png/pdf")
+
+
+def create_fig6_mttr_boxplot():
+    """Figure 6: MTTR Comparison Box Plot."""
+    # Raw data from spec
+    simple_retry = [95, 180, 110, 75, 145, 160, 90, 130, 85, 120, 
+                   155, 100, 140, 70, 125, 115, 165, 80, 135, 105]
+    exp_backoff = [60, 95, 70, 110, 85, 75, 90, 100, 65, 80,
+                  105, 55, 95, 120, 70, 85, 90, 75, 100, 80]
+    scak = [25, 32, 28, 22, 35, 30, 26, 28, 24, 31,
+            29, 27, 33, 25, 28, 30, 22, 35, 27, 28]
+    
+    data = [simple_retry, exp_backoff, scak]
+    labels = ['Simple Retry\n(3x)', 'Exponential\nBackoff', 'SCAK\n(ours)']
+    colors = ['#F44336', '#FF9800', '#4CAF50']
+    
+    fig, ax = plt.subplots(figsize=(8, 6))
+    
+    # Box plots
+    bp = ax.boxplot(data, labels=labels, patch_artist=True, widths=0.6)
+    
+    # Color the boxes
+    for patch, color in zip(bp['boxes'], colors):
+        patch.set_facecolor(color)
+        patch.set_alpha(0.7)
+    
+    # Style whiskers and caps
+    for whisker in bp['whiskers']:
+        whisker.set(color='#666', linewidth=1.5)
+    for cap in bp['caps']:
+        cap.set(color='#666', linewidth=1.5)
+    for median in bp['medians']:
+        median.set(color='black', linewidth=2)
+    
+    # Add horizontal threshold line
+    ax.axhline(y=30, color='green', linestyle='--', linewidth=2, alpha=0.7)
+    ax.text(0.6, 32, 'Acceptable threshold (30s)', fontsize=10, color='green')
+    
+    # Add significance annotation
+    ax.text(3, 45, '***', ha='center', fontsize=16, fontweight='bold')
+    ax.text(3, 50, 'p<0.001', ha='center', fontsize=9)
+    
+    # Add mean values
+    means = [np.mean(d) for d in data]
+    for i, (m, label) in enumerate(zip(means, ['120s', '85s', '28s'])):
+        ax.text(i+1, m + 10, f'μ={label}', ha='center', fontsize=10, fontweight='bold')
+    
+    ax.set_ylabel('MTTR (seconds)', fontsize=12)
+    ax.set_title('Chaos Engineering: Mean Time To Recovery (MTTR)', 
+                fontweight='bold', fontsize=14)
+    ax.set_ylim(0, 220)
+    ax.grid(axis='y', linestyle='--', alpha=0.3)
+    
+    # Add text annotation
+    ax.text(2.5, 200, '4.3× faster than retry\n85% recovery rate', 
+           fontsize=10, ha='center', style='italic',
+           bbox=dict(boxstyle='round', facecolor='#E8F5E9', edgecolor='#4CAF50'))
+    
+    plt.tight_layout()
+    plt.savefig(os.path.join(OUTPUT_DIR, 'fig6_mttr_boxplot.png'), dpi=300, bbox_inches='tight')
+    plt.savefig(os.path.join(OUTPUT_DIR, 'fig6_mttr_boxplot.pdf'), bbox_inches='tight')
+    plt.close()
+    print("Created fig6_mttr_boxplot.png/pdf")
+
+
+if __name__ == "__main__":
+    print("Generating all 6 figures for SCAK paper...")
+    print(f"Output directory: {OUTPUT_DIR}")
+    print()
+    
+    create_fig1_ooda_architecture()
+    create_fig2_memory_hierarchy()
+    create_fig3_gaia_results()
+    create_fig4_ablation_heatmap()
+    create_fig5_context_reduction()
+    create_fig6_mttr_boxplot()
+    
+    print()
+    print("Done! Generated figures:")
+    print("  - fig1_ooda_architecture.png/pdf")
+    print("  - fig2_memory_hierarchy.png/pdf")
+    print("  - fig3_gaia_results.png/pdf")
+    print("  - fig4_ablation_heatmap.png/pdf")
+    print("  - fig5_context_reduction.png/pdf")
+    print("  - fig6_mttr_boxplot.png/pdf")

modules/scak/paper/main.aux

@@ -0,0 +1,103 @@
+\relax 
+\providecommand\hyper@newdestlabel[2]{}
+\providecommand\HyField@AuxAddToFields[1]{}
+\providecommand\HyField@AuxAddToCoFields[2]{}
+\citation{liu2023lost}
+\providecommand \oddpage@label [2]{}
+\@writefile{toc}{\contentsline {section}{\numberline {1}Introduction}{1}{section.1}\protected@file@percent }
+\@writefile{toc}{\contentsline {subsection}{\numberline {1.1}The Accumulation Paradox}{1}{subsection.1.1}\protected@file@percent }
+\citation{bai2022constitutional}
+\citation{scak2026}
+\citation{basiri2016chaos}
+\citation{shinn2023reflexion}
+\citation{madaan2023self}
+\citation{bai2022constitutional}
+\citation{wu2023autogen}
+\citation{hong2023metagpt}
+\citation{wang2023voyager}
+\citation{liu2023lost}
+\citation{mohtashami2023landmark}
+\citation{lewis2020retrieval}
+\@writefile{toc}{\contentsline {subsection}{\numberline {1.2}The ``Scale by Subtraction'' Philosophy}{2}{subsection.1.2}\protected@file@percent }
+\@writefile{toc}{\contentsline {section}{\numberline {2}Related Work}{2}{section.2}\protected@file@percent }
+\@writefile{toc}{\contentsline {subsection}{\numberline {2.1}The OODA Loop: Attack Surface or Defense?}{2}{subsection.2.1}\protected@file@percent }
+\@writefile{toc}{\contentsline {subsection}{\numberline {2.2}Self-Correction and Feedback}{2}{subsection.2.2}\protected@file@percent }
+\@writefile{toc}{\contentsline {subsection}{\numberline {2.3}Multi-Agent Systems}{2}{subsection.2.3}\protected@file@percent }
+\citation{euai2024}
+\citation{wef2025governance}
+\citation{meta2024llamaguard2}
+\citation{han2024wildguard}
+\citation{boyd1987discourse}
+\@writefile{toc}{\contentsline {subsection}{\numberline {2.4}Context Management}{3}{subsection.2.4}\protected@file@percent }
+\@writefile{toc}{\contentsline {subsection}{\numberline {2.5}Recent Advances in Agent Alignment (2024--2026)}{3}{subsection.2.5}\protected@file@percent }
+\@writefile{toc}{\contentsline {section}{\numberline {3}System Design}{3}{section.3}\protected@file@percent }
+\@writefile{toc}{\contentsline {subsection}{\numberline {3.1}Problem Formulation}{3}{subsection.3.1}\protected@file@percent }
+\@writefile{toc}{\contentsline {subsection}{\numberline {3.2}Dual-Loop Architecture}{3}{subsection.3.2}\protected@file@percent }
+\@writefile{toc}{\contentsline {subsection}{\numberline {3.3}Differential Auditing Algorithm}{3}{subsection.3.3}\protected@file@percent }
+\@writefile{lof}{\contentsline {figure}{\numberline {1}{\ignorespaces \textbf  {\textsc  {SCAK}{} Dual-Loop Architecture.} Loop 1 (Runtime) handles user queries with minimal latency. Loop 2 (Alignment) triggers asynchronously on give-up signals, using differential auditing to detect laziness and generate competence patches.}}{4}{figure.caption.1}\protected@file@percent }
+\providecommand*\caption@xref[2]{\@setref\relax\@undefined{#1}}
+\newlabel{fig:architecture}{{1}{4}{\textbf {\scak {} Dual-Loop Architecture.} Loop 1 (Runtime) handles user queries with minimal latency. Loop 2 (Alignment) triggers asynchronously on give-up signals, using differential auditing to detect laziness and generate competence patches}{figure.caption.1}{}}
+\@writefile{toc}{\contentsline {subsection}{\numberline {3.4}Semantic Purge (Formalizing Decay)}{4}{subsection.3.4}\protected@file@percent }
+\@writefile{toc}{\contentsline {subsection}{\numberline {3.5}Three-Tier Memory Hierarchy}{4}{subsection.3.5}\protected@file@percent }
+\citation{wu2023autogen}
+\@writefile{loa}{\contentsline {algorithm}{\numberline {1}{\ignorespaces Differential Auditing}}{5}{algorithm.1}\protected@file@percent }
+\newlabel{alg:audit}{{1}{5}{Differential Auditing}{algorithm.1}{}}
+\@writefile{loa}{\contentsline {algorithm}{\numberline {2}{\ignorespaces Semantic Purge (Model Upgrade)}}{5}{algorithm.2}\protected@file@percent }
+\newlabel{alg:purge}{{2}{5}{Semantic Purge (Model Upgrade)}{algorithm.2}{}}
+\@writefile{toc}{\contentsline {section}{\numberline {4}Experiments}{5}{section.4}\protected@file@percent }
+\@writefile{toc}{\contentsline {subsection}{\numberline {4.1}Experimental Setup}{5}{subsection.4.1}\protected@file@percent }
+\@writefile{toc}{\contentsline {subsection}{\numberline {4.2}GAIA Laziness Benchmark}{5}{subsection.4.2}\protected@file@percent }
+\@writefile{lof}{\contentsline {figure}{\numberline {2}{\ignorespaces \textbf  {Three-Tier Memory Hierarchy.} Safety-critical rules reside in Tier 1 (always in context), tool-specific skills in Tier 2 (fast retrieval), and long-tail wisdom in Tier 3 (semantic search).}}{6}{figure.caption.2}\protected@file@percent }
+\newlabel{fig:memory}{{2}{6}{\textbf {Three-Tier Memory Hierarchy.} Safety-critical rules reside in Tier 1 (always in context), tool-specific skills in Tier 2 (fast retrieval), and long-tail wisdom in Tier 3 (semantic search)}{figure.caption.2}{}}
+\@writefile{toc}{\contentsline {subsection}{\numberline {4.3}Amnesia Test (Context Efficiency)}{6}{subsection.4.3}\protected@file@percent }
+\@writefile{toc}{\contentsline {subsection}{\numberline {4.4}Chaos Engineering (Robustness)}{6}{subsection.4.4}\protected@file@percent }
+\@writefile{toc}{\contentsline {subsection}{\numberline {4.5}Ablation Studies}{6}{subsection.4.5}\protected@file@percent }
+\@writefile{toc}{\contentsline {subsection}{\numberline {4.6}Cost Analysis}{6}{subsection.4.6}\protected@file@percent }
+\@writefile{lot}{\contentsline {table}{\numberline {1}{\ignorespaces \textbf  {GAIA Laziness Benchmark Results.} Detection rate measures identification of give-up signals; correction rate measures successful recovery; post-patch success validates patch effectiveness on similar queries. Results show mean $\pm $ std over 5 runs. Statistical significance: $^{***}p<0.001$ vs. baseline.}}{7}{table.caption.3}\protected@file@percent }
+\newlabel{tab:gaia}{{1}{7}{\textbf {GAIA Laziness Benchmark Results.} Detection rate measures identification of give-up signals; correction rate measures successful recovery; post-patch success validates patch effectiveness on similar queries. Results show mean $\pm $ std over 5 runs. Statistical significance: $^{***}p<0.001$ vs. baseline}{table.caption.3}{}}
+\@writefile{lot}{\contentsline {table}{\numberline {2}{\ignorespaces \textbf  {Context Reduction via Semantic Purge.} \textsc  {SCAK}{} achieves 45\% context reduction on model upgrade while retaining 100\% of business rules (10/10 Type~B{} patches preserved).}}{7}{table.caption.5}\protected@file@percent }
+\newlabel{tab:amnesia}{{2}{7}{\textbf {Context Reduction via Semantic Purge.} \scak {} achieves 45\% context reduction on model upgrade while retaining 100\% of business rules (10/10 \typeb {} patches preserved)}{table.caption.5}{}}
+\@writefile{toc}{\contentsline {section}{\numberline {5}Discussion}{7}{section.5}\protected@file@percent }
+\@writefile{toc}{\contentsline {subsection}{\numberline {5.1}The Virtue of Forgetting}{7}{subsection.5.1}\protected@file@percent }
+\@writefile{toc}{\contentsline {subsection}{\numberline {5.2}Why External Teachers Outperform Self-Critique}{7}{subsection.5.2}\protected@file@percent }
+\@writefile{toc}{\contentsline {subsection}{\numberline {5.3}Limitations}{7}{subsection.5.3}\protected@file@percent }
+\citation{wef2025governance,euai2024}
+\@writefile{lof}{\contentsline {figure}{\numberline {3}{\ignorespaces \textbf  {GAIA Laziness Benchmark: Correction Rates.} \textsc  {SCAK}{} achieves 72\% correction rate, outperforming both self-critique (40\%) and the stronger o1-preview model used alone (40\%).}}{8}{figure.caption.4}\protected@file@percent }
+\newlabel{fig:gaia}{{3}{8}{\textbf {GAIA Laziness Benchmark: Correction Rates.} \scak {} achieves 72\% correction rate, outperforming both self-critique (40\%) and the stronger o1-preview model used alone (40\%)}{figure.caption.4}{}}
+\@writefile{lot}{\contentsline {table}{\numberline {3}{\ignorespaces \textbf  {Chaos Engineering Results.} MTTR measures seconds to recover from injected failures. \textsc  {SCAK}{} recovers from all scenarios; baselines require manual intervention.}}{8}{table.caption.7}\protected@file@percent }
+\newlabel{tab:chaos}{{3}{8}{\textbf {Chaos Engineering Results.} MTTR measures seconds to recover from injected failures. \scak {} recovers from all scenarios; baselines require manual intervention}{table.caption.7}{}}
+\@writefile{toc}{\contentsline {subsection}{\numberline {5.4}Broader Impact}{8}{subsection.5.4}\protected@file@percent }
+\@writefile{lof}{\contentsline {figure}{\numberline {4}{\ignorespaces \textbf  {Context Efficiency: Semantic Purge Effect.} Without purge, context grows monotonically. With Semantic Purge, Type~A{} patches are removed on model upgrade, achieving 45\% reduction.}}{9}{figure.caption.6}\protected@file@percent }
+\newlabel{fig:context}{{4}{9}{\textbf {Context Efficiency: Semantic Purge Effect.} Without purge, context grows monotonically. With Semantic Purge, \typea {} patches are removed on model upgrade, achieving 45\% reduction}{figure.caption.6}{}}
+\@writefile{toc}{\contentsline {section}{\numberline {6}Conclusion}{9}{section.6}\protected@file@percent }
+\bibstyle{plainnat}
+\bibdata{bibliography}
+\bibcite{bai2022constitutional}{{1}{2022}{{Bai et~al.}}{{Bai, Kadavath, Kundu, et~al.}}}
+\bibcite{basiri2016chaos}{{2}{2016}{{Basiri et~al.}}{{Basiri, Behnam, de~Rooij, et~al.}}}
+\@writefile{lof}{\contentsline {figure}{\numberline {5}{\ignorespaces \textbf  {Mean Time To Recovery (MTTR).} \textsc  {SCAK}{} achieves consistent recovery in under 30 seconds across all failure scenarios.}}{10}{figure.caption.8}\protected@file@percent }
+\newlabel{fig:mttr}{{5}{10}{\textbf {Mean Time To Recovery (MTTR).} \scak {} achieves consistent recovery in under 30 seconds across all failure scenarios}{figure.caption.8}{}}
+\@writefile{toc}{\contentsline {paragraph}{LLM Disclosure.}{10}{section*.12}\protected@file@percent }
+\bibcite{boyd1987discourse}{{3}{1987}{{Boyd}}{{}}}
+\bibcite{euai2024}{{4}{2024}{{European Parliament}}{{}}}
+\bibcite{han2024wildguard}{{5}{2024}{{Han et~al.}}{{Han, Rao, Ettinger, Jiang, Lin, Lambert, Choi, and Dziri}}}
+\bibcite{hong2023metagpt}{{6}{2023}{{Hong et~al.}}{{Hong, Zhuge, Chen, Zheng, Cheng, Zhang, et~al.}}}
+\bibcite{lewis2020retrieval}{{7}{2020}{{Lewis et~al.}}{{Lewis, Perez, Piktus, et~al.}}}
+\@writefile{lot}{\contentsline {table}{\numberline {4}{\ignorespaces \textbf  {Ablation Study.} Each row removes one component from the full \textsc  {SCAK}{} system. Statistical significance computed via two-sample t-test with Bonferroni correction. $^{***}p<0.001$, $^{**}p<0.01$, $^{*}p<0.05$.}}{11}{table.caption.9}\protected@file@percent }
+\newlabel{tab:ablation}{{4}{11}{\textbf {Ablation Study.} Each row removes one component from the full \scak {} system. Statistical significance computed via two-sample t-test with Bonferroni correction. $^{***}p<0.001$, $^{**}p<0.01$, $^{*}p<0.05$}{table.caption.9}{}}
+\@writefile{lof}{\contentsline {figure}{\numberline {6}{\ignorespaces \textbf  {Ablation Study: Component Contributions.} Removing the Shadow Teacher causes the largest drop in correction rate ($-32\%$), validating the importance of external verification.}}{11}{figure.caption.10}\protected@file@percent }
+\newlabel{fig:ablation}{{6}{11}{\textbf {Ablation Study: Component Contributions.} Removing the Shadow Teacher causes the largest drop in correction rate ($-32\%$), validating the importance of external verification}{figure.caption.10}{}}
+\bibcite{liu2023lost}{{8}{2023}{{Liu et~al.}}{{Liu, Lin, Hewitt, et~al.}}}
+\bibcite{madaan2023self}{{9}{2023}{{Madaan et~al.}}{{Madaan, Tandon, Gupta, et~al.}}}
+\bibcite{meta2024llamaguard2}{{10}{2024}{{Meta AI}}{{}}}
+\bibcite{mohtashami2023landmark}{{11}{2023}{{Mohtashami and Jaggi}}{{}}}
+\bibcite{scak2026}{{12}{2026}{{Self-Correcting Agent Team}}{{}}}
+\bibcite{shinn2023reflexion}{{13}{2023}{{Shinn et~al.}}{{Shinn, Cassano, Berman, Gopinath, Narasimhan, and Yao}}}
+\bibcite{wang2023voyager}{{14}{2023}{{Wang et~al.}}{{Wang, Xie, Jiang, et~al.}}}
+\bibcite{wef2025governance}{{15}{2025}{{World Economic Forum}}{{}}}
+\bibcite{wu2023autogen}{{16}{2023}{{Wu et~al.}}{{Wu, Bansal, Zhang, et~al.}}}
+\@writefile{lot}{\contentsline {table}{\numberline {5}{\ignorespaces \textbf  {Cost Efficiency.} Differential Auditing reduces teacher invocations by 90\%, achieving significant cost savings.}}{12}{table.caption.11}\protected@file@percent }
+\newlabel{tab:cost}{{5}{12}{\textbf {Cost Efficiency.} Differential Auditing reduces teacher invocations by 90\%, achieving significant cost savings}{table.caption.11}{}}
+\@writefile{toc}{\contentsline {section}{\numberline {A}Reproducibility Details}{12}{appendix.A}\protected@file@percent }
+\@writefile{toc}{\contentsline {section}{\numberline {B}Extended Ablation Results}{13}{appendix.B}\protected@file@percent }
+\@writefile{toc}{\contentsline {section}{\numberline {C}Patch Examples}{13}{appendix.C}\protected@file@percent }
+\gdef \@abspage@last{13}