Add files using upload-large-folder tool

agg_coarse/all/sweep_component_comparison/05_circuit_synthesis/circuit_summary.txt

@@ -0,0 +1,78 @@
+# Circuit Summary
+
+This document summarizes the inferred circuit from activation patching analysis.
+Scores shown as: noising_disruption / denoising_recovery
+
+## Key Finding: Two-Phase Architecture
+
+The circuit operates in two phases:
+1. **Attention phase**: Attention heads read from source positions and route information
+2. **MLP phase**: MLP layers transform and write to destination positions
+
+## Source Positions (Early)
+
+Found 4 key source positions (1 have backup pathways):
+  - P83: 0.963/0.016 (important)
+  - P78: 0.787/-0.049 (important)
+  - P73: 0.752/0.931 (REDUNDANT - has backups)
+  - P98: 0.053/-0.001 (important)
+
+Interpretation: Redundant positions have denoising > noising, meaning
+the model can recover even when these positions are corrupted (backup pathways exist).
+
+## Attention Layers
+
+Top attention layers (ranked by noising disruption):
+  - L24: 0.414/0.364 ★ MULTI-FUNCTION
+  - L30: 0.165/0.152
+  - L29: 0.162/0.137
+  - L23: 0.157/0.198
+  - L21: 0.153/0.209
+
+These layers read from source positions and route information.
+Multi-function layers (L24) appear in both attention AND MLP top lists,
+suggesting they do double-duty processing.
+
+## MLP Layers
+
+Top MLP layers (ranked by noising disruption):
+  - L35: 0.262/0.266
+  - L31: 0.208/0.174
+  - L24: 0.151/0.208 ★ MULTI-FUNCTION
+  - L34: 0.148/0.095
+  - L27: 0.088/0.064
+
+These layers transform information and write to destination positions.
+
+## Destination Positions (Late) - THE BOTTLENECK
+
+Found 4 key destination positions:
+  - P138: 1.000/0.000 (★ BOTTLENECK)
+  - P118: 0.186/0.000 (secondary)
+  - P133: 0.171/0.000 (secondary)
+  - P103: 0.024/-0.001 (secondary)
+
+CRITICAL: 1 position(s) are true bottlenecks (noising > 0.5).
+Corrupting these positions severely disrupts the model's output.
+This is where the final answer is computed.
+
+## Layer-Position Bindings
+
+These show which layers act at which positions (the actual circuit edges):
+
+  - L21-L30 attention reads from P73-P98
+  - L24-L35 MLP writes to P103-P138
+
+## Methodology Notes
+
+- **Noising** (disruption): Replace clean activation with corrupted → measures necessity
+- **Denoising** (recovery): Replace corrupted activation with clean → measures sufficiency
+- High noising + low denoising = necessary but has backups (OR relationship)
+- High noising + high denoising = necessary AND sufficient (AND relationship)
+- The redundancy gap (noising - denoising) indicates how replaceable a component is
+- **Denoising-only positions filtered**: Positions where denoising >> noising are excluded
+  as they indicate artifacts rather than true circuit components
+
+## Circuit Diagram
+
+See information_flow_diagram.png for a visual representation.

agg_coarse/all/sweep_component_comparison/README.md

@@ -0,0 +1,104 @@
+# Component Comparison Plots
+
+This directory contains activation patching analysis visualizations organized
+into five categories. Each plot has a figure number for easy reference.
+
+## Quick Start
+
+1. **Check 01_sanity_checks/** first - if these fail, results are unreliable
+2. **Read 05_circuit_synthesis/circuit_summary.txt** for the plain-text explanation
+3. **View 05_circuit_synthesis/information_flow_diagram.png** for the visual summary
+4. Drill into other folders for detailed analysis
+
+---
+
+## Directory Structure
+
+### 01_sanity_checks/
+*Validation plots - check these first. If they fail, everything else is suspect.*
+
+| Fig | File | Description |
+|-----|------|-------------|
+| 1.1 | resid_delta_sanity.png | Compares resid_pre[L+1] vs resid_post[L] (should match) |
+| 1.2 | resid_delta_difference.png | Difference plot with +/-0.02 tolerance band |
+
+### 02_overview/
+*The big picture: "Where does the computation happen?"*
+
+| Fig | File | Description |
+|-----|------|-------------|
+| 2.1 | heatmap_layers_denoising.png | Layer × Component attribution (denoising recovery) |
+| 2.2 | heatmap_layers_noising.png | Layer × Component attribution (noising disruption) |
+| 2.3 | heatmap_layers_colnorm_denoising.png | Column-normalized heatmap (reveals fine structure) |
+| 2.4 | heatmap_layers_colnorm_noising.png | Column-normalized heatmap (noising) |
+| 2.5 | heatmap_positions_denoising.png | Position × Component attribution (denoising) |
+| 2.6 | heatmap_positions_noising.png | Position × Component attribution (noising) |
+| 2.7 | layer_position_heatmap.png | Layer × Position importance map (2D localization) |
+
+### 03_component_decomp/
+*Drilling in: "Is it attention or MLP? Which specific components?"*
+
+| Fig | File | Description |
+|-----|------|-------------|
+| 3.1 | attn_vs_mlp_layer.png | Attention vs MLP scatter by layer |
+| 3.2 | attn_vs_mlp_position.png | Attention vs MLP scatter by position |
+| 3.3 | attn_vs_mlp_paired.png | Paired scatter with arrows (denoising→noising movement) |
+| 3.4 | component_importance_ranked.png | Top components ranked with same-layer linking |
+| 3.5 | cumulative_recovery.png | Cumulative recovery with dip annotations |
+| 3.6 | marginal_contribution.png | Per-layer marginal contribution |
+| 3.7 | position_component_interaction.png | Effect vs position for each component |
+| 3.8 | position_interaction_zoomed.png | Zoomed panels for hub regions |
+
+### 04_redundancy/
+*Methodology comparison: "Can I trust the scores? What has backup pathways?"*
+
+| Fig | File | Description |
+|-----|------|-------------|
+| 4.1 | noise_vs_denoise_per_component_layer.png | Noising vs denoising scatter (layers) |
+| 4.2 | noise_vs_denoise_per_component_position.png | Noising vs denoising scatter (positions) |
+| 4.3 | redundancy_gap.png | Redundancy gap (disruption - recovery) per layer |
+| 4.4 | redundancy_gap_sorted.png | Redundancy gap sorted by magnitude |
+| 4.5 | difference_heatmap_layer.png | Redundancy gap heatmap (layers) |
+| 4.6 | difference_heatmap_position.png | Redundancy gap heatmap (positions) |
+
+### 05_circuit_synthesis/
+*The punchline: synthesized information flow.*
+
+| Fig | File | Description |
+|-----|------|-------------|
+| 5.1 | information_flow_diagram.png | Circuit summary with redundancy and bottleneck annotations |
+| — | circuit_summary.txt | Plain-text explanation of the circuit hypothesis |
+
+---
+
+## Interpretation Guide
+
+### Scatter Plot Regions
+- **AND region** (upper-right): High in both → necessary AND sufficient
+- **OR region** (lower-right): High denoising, low noising → has backup pathways
+
+### Redundancy Gap
+- **Positive gap** (disruption > recovery): Component is necessary
+- **Negative gap** (recovery > disruption): Component is sufficient but replaceable
+
+### Scores in Circuit Diagram
+- Format: noising_disruption / denoising_recovery
+- Higher noising = more necessary (corrupting it breaks the model)
+- Higher denoising = more sufficient (restoring it fixes the model)
+
+### Visual Conventions
+- ★ = Multi-function layer (appears in both attention and MLP top lists)
+- Colored brackets = Same-layer components (e.g., L24_attn and L24_mlp)
+- Red thick arrow = Critical bottleneck path
+- Dashed arrow = Backup/bypass pathway
+
+---
+
+## Colormap Reference
+
+| Plot Type | Colormap | Interpretation |
+|-----------|----------|----------------|
+| Recovery/Disruption heatmaps | RdYlGn | Green=high, Red=low |
+| 2D localization map | Hot | White=high, Black=low |
+| Redundancy gap heatmaps | RdBu_r | Red=necessity, Blue=sufficiency |
+| Layer/position scatter | Viridis | Yellow=late, Purple=early |

agg_coarse/horizon/sweep_component_comparison/README.md

@@ -0,0 +1,104 @@
+# Component Comparison Plots
+
+This directory contains activation patching analysis visualizations organized
+into five categories. Each plot has a figure number for easy reference.
+
+## Quick Start
+
+1. **Check 01_sanity_checks/** first - if these fail, results are unreliable
+2. **Read 05_circuit_synthesis/circuit_summary.txt** for the plain-text explanation
+3. **View 05_circuit_synthesis/information_flow_diagram.png** for the visual summary
+4. Drill into other folders for detailed analysis
+
+---
+
+## Directory Structure
+
+### 01_sanity_checks/
+*Validation plots - check these first. If they fail, everything else is suspect.*
+
+| Fig | File | Description |
+|-----|------|-------------|
+| 1.1 | resid_delta_sanity.png | Compares resid_pre[L+1] vs resid_post[L] (should match) |
+| 1.2 | resid_delta_difference.png | Difference plot with +/-0.02 tolerance band |
+
+### 02_overview/
+*The big picture: "Where does the computation happen?"*
+
+| Fig | File | Description |
+|-----|------|-------------|
+| 2.1 | heatmap_layers_denoising.png | Layer × Component attribution (denoising recovery) |
+| 2.2 | heatmap_layers_noising.png | Layer × Component attribution (noising disruption) |
+| 2.3 | heatmap_layers_colnorm_denoising.png | Column-normalized heatmap (reveals fine structure) |
+| 2.4 | heatmap_layers_colnorm_noising.png | Column-normalized heatmap (noising) |
+| 2.5 | heatmap_positions_denoising.png | Position × Component attribution (denoising) |
+| 2.6 | heatmap_positions_noising.png | Position × Component attribution (noising) |
+| 2.7 | layer_position_heatmap.png | Layer × Position importance map (2D localization) |
+
+### 03_component_decomp/
+*Drilling in: "Is it attention or MLP? Which specific components?"*
+
+| Fig | File | Description |
+|-----|------|-------------|
+| 3.1 | attn_vs_mlp_layer.png | Attention vs MLP scatter by layer |
+| 3.2 | attn_vs_mlp_position.png | Attention vs MLP scatter by position |
+| 3.3 | attn_vs_mlp_paired.png | Paired scatter with arrows (denoising→noising movement) |
+| 3.4 | component_importance_ranked.png | Top components ranked with same-layer linking |
+| 3.5 | cumulative_recovery.png | Cumulative recovery with dip annotations |
+| 3.6 | marginal_contribution.png | Per-layer marginal contribution |
+| 3.7 | position_component_interaction.png | Effect vs position for each component |
+| 3.8 | position_interaction_zoomed.png | Zoomed panels for hub regions |
+
+### 04_redundancy/
+*Methodology comparison: "Can I trust the scores? What has backup pathways?"*
+
+| Fig | File | Description |
+|-----|------|-------------|
+| 4.1 | noise_vs_denoise_per_component_layer.png | Noising vs denoising scatter (layers) |
+| 4.2 | noise_vs_denoise_per_component_position.png | Noising vs denoising scatter (positions) |
+| 4.3 | redundancy_gap.png | Redundancy gap (disruption - recovery) per layer |
+| 4.4 | redundancy_gap_sorted.png | Redundancy gap sorted by magnitude |
+| 4.5 | difference_heatmap_layer.png | Redundancy gap heatmap (layers) |
+| 4.6 | difference_heatmap_position.png | Redundancy gap heatmap (positions) |
+
+### 05_circuit_synthesis/
+*The punchline: synthesized information flow.*
+
+| Fig | File | Description |
+|-----|------|-------------|
+| 5.1 | information_flow_diagram.png | Circuit summary with redundancy and bottleneck annotations |
+| — | circuit_summary.txt | Plain-text explanation of the circuit hypothesis |
+
+---
+
+## Interpretation Guide
+
+### Scatter Plot Regions
+- **AND region** (upper-right): High in both → necessary AND sufficient
+- **OR region** (lower-right): High denoising, low noising → has backup pathways
+
+### Redundancy Gap
+- **Positive gap** (disruption > recovery): Component is necessary
+- **Negative gap** (recovery > disruption): Component is sufficient but replaceable
+
+### Scores in Circuit Diagram
+- Format: noising_disruption / denoising_recovery
+- Higher noising = more necessary (corrupting it breaks the model)
+- Higher denoising = more sufficient (restoring it fixes the model)
+
+### Visual Conventions
+- ★ = Multi-function layer (appears in both attention and MLP top lists)
+- Colored brackets = Same-layer components (e.g., L24_attn and L24_mlp)
+- Red thick arrow = Critical bottleneck path
+- Dashed arrow = Backup/bypass pathway
+
+---
+
+## Colormap Reference
+
+| Plot Type | Colormap | Interpretation |
+|-----------|----------|----------------|
+| Recovery/Disruption heatmaps | RdYlGn | Green=high, Red=low |
+| 2D localization map | Hot | White=high, Black=low |
+| Redundancy gap heatmaps | RdBu_r | Red=necessity, Blue=sufficiency |
+| Layer/position scatter | Viridis | Yellow=late, Purple=early |

different_label.json

@@ -0,0 +1 @@
+[]

different_order.json

@@ -0,0 +1 @@
+[]

half_horizon.json

@@ -0,0 +1,16 @@
+{
+    "clean_has_horizon": [
+        0,
+        3,
+        7,
+        10,
+        13,
+        15,
+        18,
+        20,
+        22,
+        23,
+        24
+    ],
+    "corrupted_has_horizon": []
+}

horizon.json

@@ -0,0 +1,21 @@
+{
+    "clean_greater": [],
+    "corrupted_greater": [
+        1,
+        2,
+        4,
+        5,
+        6,
+        9,
+        11,
+        12,
+        14,
+        16,
+        17,
+        19
+    ],
+    "equal": [
+        8,
+        21
+    ]
+}

horizon_analysis.json

@@ -0,0 +1,42 @@
+{
+    "horizon": {
+        "clean_greater": [],
+        "corrupted_greater": [
+            1,
+            2,
+            4,
+            5,
+            6,
+            9,
+            11,
+            12,
+            14,
+            16,
+            17,
+            19
+        ],
+        "equal": [
+            8,
+            21
+        ]
+    },
+    "half_horizon": {
+        "clean_has_horizon": [
+            0,
+            3,
+            7,
+            10,
+            13,
+            15,
+            18,
+            20,
+            22,
+            23,
+            24
+        ],
+        "corrupted_has_horizon": []
+    },
+    "no_horizon": {
+        "pair_indices": []
+    }
+}

no_horizon.json

@@ -0,0 +1,3 @@
+{
+    "pair_indices": []
+}

pair_analysis.json

@@ -0,0 +1,90 @@
+{
+    "labels": {
+        "same_labels": [
+            0,
+            1,
+            2,
+            3,
+            4,
+            5,
+            6,
+            7,
+            8,
+            9,
+            10,
+            11,
+            12,
+            13,
+            14,
+            15,
+            16,
+            17,
+            18,
+            19,
+            20,
+            21,
+            22,
+            23,
+            24
+        ],
+        "different_labels": []
+    },
+    "order": {
+        "same_order": [
+            0,
+            1,
+            2,
+            3,
+            4,
+            5,
+            6,
+            7,
+            8,
+            9,
+            10,
+            11,
+            12,
+            13,
+            14,
+            15,
+            16,
+            17,
+            18,
+            19,
+            20,
+            21,
+            22,
+            23,
+            24
+        ],
+        "different_order": [],
+        "both_short_first": [
+            0,
+            1,
+            2,
+            3,
+            4,
+            5,
+            6,
+            7,
+            8,
+            9,
+            10,
+            11,
+            12,
+            13,
+            14,
+            15,
+            16,
+            17,
+            18,
+            19,
+            20,
+            21,
+            22,
+            23,
+            24
+        ],
+        "both_long_first": []
+    }
+}

profile.json

@@ -0,0 +1,185 @@
+{
+  "entries": [
+    {
+      "name": "run_experiment",
+      "total_ms": 1225781.5171249968,
+      "count": 1,
+      "avg_ms": 1225781.5171249968,
+      "children": [
+        {
+          "name": "step_preference_data",
+          "total_ms": 3706.3284589967225,
+          "count": 1,
+          "avg_ms": 3706.3284589967225
+        },
+        {
+          "name": "step_attribution_patching",
+          "total_ms": 6613.111875019968,
+          "count": 1,
+          "avg_ms": 6613.111875019968
+        },
+        {
+          "name": "step_coarse_activation_patching",
+          "total_ms": 2332.7152080019005,
+          "count": 1,
+          "avg_ms": 2332.7152080019005
+        },
+        {
+          "name": "step_diffmeans",
+          "total_ms": 283.85824998258613,
+          "count": 1,
+          "avg_ms": 283.85824998258613
+        },
+        {
+          "name": "step_geo",
+          "total_ms": 10649.801999999909,
+          "count": 1,
+          "avg_ms": 10649.801999999909
+        },
+        {
+          "name": "step_process_results",
+          "total_ms": 14.578208996681497,
+          "count": 1,
+          "avg_ms": 14.578208996681497
+        },
+        {
+          "name": "step_visualize_results",
+          "total_ms": 1202180.1247920084,
+          "count": 1,
+          "avg_ms": 1202180.1247920084,
+          "children": [
+            {
+              "name": "visualize_all_att_aggregated_slices",
+              "total_ms": 44276.10679101781,
+              "count": 1,
+              "avg_ms": 44276.10679101781,
+              "children": [
+                {
+                  "name": "visualize_att_aggregated",
+                  "total_ms": 44275.95533296699,
+                  "count": 4,
+                  "avg_ms": 11068.988833241747,
+                  "children": [
+                    {
+                      "name": "generate_analysis_plots",
+                      "total_ms": 12549.135708017275,
+                      "count": 4,
+                      "avg_ms": 3137.2839270043187,
+                      "children": [
+                        {
+                          "name": "plot_aggregated_layer_attribution",
+                          "total_ms": 3622.4987510067876,
+                          "count": 4,
+                          "avg_ms": 905.6246877516969,
+                          "children": [
+                            {
+                              "name": "plot_layer_attribution_line",
+                              "total_ms": 3606.1577108921483,
+                              "count": 32,
+                              "avg_ms": 112.69242846537963
+                            }
+                          ]
+                        },
+                        {
+                          "name": "plot_aggregated_histograms",
+                          "total_ms": 731.5591250080615,
+                          "count": 4,
+                          "avg_ms": 182.88978125201538,
+                          "children": [
+                            {
+                              "name": "plot_score_histogram",
+                              "total_ms": 730.164539971156,
+                              "count": 8,
+                              "avg_ms": 91.2705674963945
+                            }
+                          ]
+                        },
+                        {
+                          "name": "plot_aggregated_mode_comparison",
+                          "total_ms": 5748.870708950562,
+                          "count": 4,
+                          "avg_ms": 1437.2176772376406,
+                          "children": [
+                            {
+                              "name": "plot_mode_bar_chart",
+                              "total_ms": 2950.436625018483,
+                              "count": 4,
+                              "avg_ms": 737.6091562546208
+                            },
+                            {
+                              "name": "plot_mode_scatter",
+                              "total_ms": 1291.1740419804119,
+                              "count": 4,
+                              "avg_ms": 322.79351049510296
+                            },
+                            {
+                              "name": "plot_component_scatter_grid",
+                              "total_ms": 1504.7882919898257,
+                              "count": 4,
+                              "avg_ms": 376.1970729974564
+                            }
+                          ]
+                        },
+                        {
+                          "name": "plot_aggregated_top_scores",
+                          "total_ms": 2441.4216660079546,
+                          "count": 4,
+                          "avg_ms": 610.3554165019887,
+                          "children": [
+                            {
+                              "name": "plot_top_scores_text",
+                              "total_ms": 394.730250001885,
+                              "count": 8,
+                              "avg_ms": 49.341281250235625
+                            },
+                            {
+                              "name": "plot_top_scores_bar",
+                              "total_ms": 2044.0846259880345,
+                              "count": 8,
+                              "avg_ms": 255.51057824850432
+                            }
+                          ]
+                        }
+                      ]
+                    }
+                  ]
+                }
+              ]
+            },
+            {
+              "name": "visualize_all_aggregated",
+              "total_ms": 146356.77679200307,
+              "count": 1,
+              "avg_ms": 146356.77679200307
+            },
+            {
+              "name": "visualize_tokenization",
+              "total_ms": 20710.865210945485,
+              "count": 25,
+              "avg_ms": 828.4346084378194
+            },
+            {
+              "name": "visualize_att_patching",
+              "total_ms": 101914.27295503672,
+              "count": 50,
+              "avg_ms": 2038.2854591007344
+            },
+            {
+              "name": "visualize_coarse_patching",
+              "total_ms": 749217.5547430525,
+              "count": 100,
+              "avg_ms": 7492.175547430525
+            },
+            {
+              "name": "visualize_component_comparison",
+              "total_ms": 87789.18679297203,
+              "count": 25,
+              "avg_ms": 3511.5674717188813
+            }
+          ]
+        }
+      ]
+    }
+  ],
+  "total_ms": 1225781.5171249968
+}

same_label.json

@@ -0,0 +1,27 @@
+[
+    0,
+    1,
+    2,
+    3,
+    4,
+    5,
+    6,
+    7,
+    8,
+    9,
+    10,
+    11,
+    12,
+    13,
+    14,
+    15,
+    16,
+    17,
+    18,
+    19,
+    20,
+    21,
+    22,
+    23,
+    24
+]

same_order.json

@@ -0,0 +1,27 @@
+[
+    0,
+    1,
+    2,
+    3,
+    4,
+    5,
+    6,
+    7,
+    8,
+    9,
+    10,
+    11,
+    12,
+    13,
+    14,
+    15,
+    16,
+    17,
+    18,
+    19,
+    20,
+    21,
+    22,
+    23,
+    24
+]