diff --git a/pyproject.toml b/pyproject.toml
index a952a1c..46b2ef6 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -62,6 +62,7 @@ allow-direct-references = true
 [tool.pytest.ini_options]
 asyncio_mode = "auto"
 addopts = ["-p", "no:opik"]
+pythonpath = ["."]
 testpaths = ["tests"]
 
 [tool.ruff]
diff --git a/scripts/classify_regimes.py b/scripts/classify_regimes.py
index 6fbc4c7..fb58ef2 100644
--- a/scripts/classify_regimes.py
+++ b/scripts/classify_regimes.py
@@ -145,6 +145,7 @@ def main() -> None:
 
     all_turn_texts: list[str] = []
     run_turns: dict[str, list[str]] = {}
+    run_metadata: dict[str, dict[str, str | int]] = {}
 
     for model_name, task_runs in grouped.items():
         for task_id, runs in task_runs.items():
@@ -152,6 +153,11 @@ def main() -> None:
                 ts = turn_texts(run, fallback_any_message=False)
                 key = f"{model_name}/{task_id}/run{run.run_index}"
                 run_turns[key] = ts
+                run_metadata[key] = {
+                    "model": model_name,
+                    "task_id": task_id,
+                    "run_index": run.run_index,
+                }
                 all_turn_texts.extend(ts)
 
     used_fallback_messages = False
@@ -159,12 +165,18 @@ def main() -> None:
         used_fallback_messages = True
         all_turn_texts = []
         run_turns = {}
+        run_metadata = {}
         for model_name, task_runs in grouped.items():
             for task_id, runs in task_runs.items():
                 for run in runs:
                     ts = turn_texts(run, fallback_any_message=True)
                     key = f"{model_name}/{task_id}/run{run.run_index}"
                     run_turns[key] = ts
+                    run_metadata[key] = {
+                        "model": model_name,
+                        "task_id": task_id,
+                        "run_index": run.run_index,
+                    }
                     all_turn_texts.extend(ts)
 
     if not all_turn_texts:
@@ -205,6 +217,7 @@ def main() -> None:
     for key, metrics in per_run.items():
         metrics["regime"] = classify(metrics, thresholds)
         metrics["turn_source"] = "any_message" if used_fallback_messages else "assistant"
+        metrics.update(run_metadata.get(key, {}))
 
     args.reports_dir.mkdir(parents=True, exist_ok=True)
     out = args.reports_dir / "regimes.json"
diff --git a/scripts/compute_debiased_dynamics.py b/scripts/compute_debiased_dynamics.py
index 6df01df..336fb47 100644
--- a/scripts/compute_debiased_dynamics.py
+++ b/scripts/compute_debiased_dynamics.py
@@ -34,9 +34,9 @@ def compute_debiased_dynamics(regimes_path, constraint_path, weights_path, topic
     model_regimes_weight_sum = defaultdict(float)
 
     for key, data in regimes.items():
-        parts = key.split("/")
-        model = parts[0]
-        task_id = parts[1] if len(parts) > 1 else parts[0]
+        key_parts = key.rsplit("/", 2)
+        model = data.get("model") or (key_parts[0] if len(key_parts) == 3 else key)
+        task_id = data.get("task_id") or (key_parts[1] if len(key_parts) == 3 else key)
 
         # Match task to topic
         matched_topic = "unknown"
diff --git a/scripts/generate_dynamical_report.py b/scripts/generate_dynamical_report.py
index 55d52e8..bdcb087 100644
--- a/scripts/generate_dynamical_report.py
+++ b/scripts/generate_dynamical_report.py
@@ -79,7 +79,7 @@ def main() -> None:
     L("")
     by_model = defaultdict(Counter)
     for key, row in regimes.items():
-        model = key.split("/")[0]
+        model = row.get("model") or key.rsplit("/", 2)[0]
         regime = row.get("regime", "unknown")
         by_model[model][regime] += 1
 
diff --git a/scripts/run_posterior_dynamics_pipeline.py b/scripts/run_posterior_dynamics_pipeline.py
index 9c433a3..7a56b9e 100644
--- a/scripts/run_posterior_dynamics_pipeline.py
+++ b/scripts/run_posterior_dynamics_pipeline.py
@@ -88,6 +88,15 @@ def main() -> None:
     _run([py, str(scripts_dir / "variance_decomp.py"), "--archive-dir", str(archive_dir), "--reports-dir", str(reports_dir), *tier_args])
     _run([py, str(scripts_dir / "survival_analysis.py"), "--archive-dir", str(archive_dir), "--reports-dir", str(reports_dir), *tier_args])
     _run([py, str(scripts_dir / "snr_weighted_ranking.py"), "--archive-dir", str(archive_dir), "--reports-dir", str(reports_dir), *tier_args])
+    _run([
+        py,
+        str(scripts_dir / "violation_time_decomposition.py"),
+        "--archive-dir",
+        str(archive_dir),
+        "--reports-dir",
+        str(reports_dir / "violation_time_decomposition"),
+        *tier_args,
+    ])
     _run([py, str(scripts_dir / "generate_dynamical_report.py"), "--reports-dir", str(reports_dir)])
     if args.include_dynamics_report:
         _write_dynamics_reports(archive_dir, output_dir, args.tier)
diff --git a/scripts/violation_time_decomposition.py b/scripts/violation_time_decomposition.py
new file mode 100644
index 0000000..6f8500a
--- /dev/null
+++ b/scripts/violation_time_decomposition.py
@@ -0,0 +1,215 @@
+#!/usr/bin/env python3
+import argparse
+import json
+import math
+import sys
+from pathlib import Path
+from collections import defaultdict
+
+sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
+from clawbench.dynamics_archive import load_task_runs_by_model
+from clawbench.trajectory import extract_shell_command, has_dangerous_shell_pattern
+
+
+def safe_name(value: str) -> str:
+    return value.replace("/", "_").replace(":", "_")
+
+def get_first_violation_turn(run):
+    """
+    Returns (turn_index, has_violation)
+    Turn index is 1-based.
+    If no violation, turn_index is the max turn + 1, has_violation is False.
+    """
+    if not run.trajectory_result.forbidden_violations:
+        return len(run.transcript.assistant_messages) + 1, False
+
+    for i, msg in enumerate(run.transcript.assistant_messages, 1):
+        for tc in msg.tool_calls:
+            command = extract_shell_command(tc)
+            if command and has_dangerous_shell_pattern(command):
+                return i, True
+    return max(len(run.transcript.assistant_messages), 1), True
+
+def compute_decomposition(events_by_topic, max_t=20):
+    """
+    events_by_topic: dict mapping topic (e.g., scenario) to list of (turn, has_violation)
+    Returns decomposition metrics.
+    """
+    # Flatten all events
+    all_events = []
+    for evs in events_by_topic.values():
+        all_events.extend(evs)
+    
+    total = len(all_events)
+    if total == 0:
+        return {}
+    
+    metrics = {
+        "marginal_hazard": [],
+        "marginal_survival": [],
+        "conditional_hazards": defaultdict(list),
+        "mutual_information": []
+    }
+    
+    # Pre-calculate marginal
+    for t in range(1, max_t + 1):
+        at_risk_total = sum(1 for tf, _ in all_events if tf >= t)
+        events_total = sum(1 for tf, is_event in all_events if is_event and tf == t)
+        survived_total = sum(1 for tf, is_event in all_events if (not is_event and tf >= t) or (is_event and tf > t))
+        
+        h_t = events_total / at_risk_total if at_risk_total > 0 else 0.0
+        s_t = survived_total / total if total > 0 else 0.0
+        
+        metrics["marginal_hazard"].append(h_t)
+        metrics["marginal_survival"].append(s_t)
+        
+        # Calculate conditional hazards
+        mi_t = 0.0
+        for topic, evs in events_by_topic.items():
+            at_risk_topic = sum(1 for tf, _ in evs if tf >= t)
+            events_topic = sum(1 for tf, is_event in evs if is_event and tf == t)
+            h_t_given_s = events_topic / at_risk_topic if at_risk_topic > 0 else 0.0
+            metrics["conditional_hazards"][topic].append(h_t_given_s)
+            
+            # P(S = topic | T >= t)
+            if at_risk_total > 0 and at_risk_topic > 0:
+                p_s_given_at_risk = at_risk_topic / at_risk_total
+                
+                # MI term: P(S|T>=t) * D_KL( P(V|S) || P(V) )
+                kl = 0.0
+                if h_t_given_s > 0 and h_t > 0:
+                    kl += h_t_given_s * math.log2(h_t_given_s / h_t)
+                if (1 - h_t_given_s) > 0 and (1 - h_t) > 0:
+                    kl += (1 - h_t_given_s) * math.log2((1 - h_t_given_s) / (1 - h_t))
+                
+                mi_t += p_s_given_at_risk * kl
+                
+        metrics["mutual_information"].append(mi_t)
+        
+    return metrics
+
+def plot_metrics(metrics, model_name, output_dir):
+    try:
+        import matplotlib
+
+        matplotlib.use("Agg")
+        import matplotlib.pyplot as plt
+    except ModuleNotFoundError:
+        return None
+
+    max_t = len(metrics["marginal_hazard"])
+    turns = list(range(1, max_t + 1))
+    
+    plt.figure(figsize=(15, 5))
+    
+    # Plot 1: Marginal Hazard
+    plt.subplot(1, 3, 1)
+    plt.plot(turns, metrics["marginal_hazard"], marker='o', color='red')
+    plt.title("Marginal Hazard $h(t)$")
+    plt.xlabel("Turn $t$")
+    plt.ylabel("P(Violation | No Prior Violation)")
+    plt.grid(True)
+    
+    # Plot 2: Conditional Hazards
+    plt.subplot(1, 3, 2)
+    for topic, h_cond in metrics["conditional_hazards"].items():
+        if max(h_cond) > 0: # Only plot if there's non-zero hazard
+            plt.plot(turns, h_cond, alpha=0.5, label=topic)
+    plt.title("Conditional Hazards $h(t | Topic)$")
+    plt.xlabel("Turn $t$")
+    # plt.legend(bbox_to_anchor=(1.05, 1), loc='upper left', fontsize='small')
+    plt.grid(True)
+    
+    # Plot 3: Mutual Information
+    plt.subplot(1, 3, 3)
+    plt.plot(turns, metrics["mutual_information"], marker='o', color='purple')
+    plt.title(r"Mutual Information $I(Topic; V_t | T \geq t)$")
+    plt.xlabel("Turn $t$")
+    plt.ylabel("Bits")
+    plt.grid(True)
+    
+    plt.tight_layout()
+    plot_path = output_dir / f"violation_metrics_{model_name}.png"
+    plt.savefig(plot_path)
+    plt.close()
+    return plot_path
+
+def generate_markdown(metrics, model_name, plot_path, out_file):
+    max_t = len(metrics["marginal_hazard"])
+    
+    md = [
+        f"# Spatio-Temporal Violation Dynamics: {model_name}",
+        "",
+        "## Theoretical Decomposition",
+        "This report decomposes the time-to-first-violation for safety triggers and tool misuses.",
+        "To connect the long-term behavior of agent risk to its spatial risk conditioned on context semantics, we decompose the first occurrence probability:",
+        r"$$ P(T = t) = h(t) \cdot S(t-1) $$",
+        "where $h(t)$ is the conditional hazard rate at turn $t$, and $S(t-1)$ is the macro survival probability.",
+        "",
+        "Furthermore, we examine the mutual information between the semantic spatial context (scenario) and the violation event at each turn to determine if localized contexts explain hazard spikes.",
+        "",
+        "## Visualization",
+        f"![Violation Metrics]({plot_path.name})" if plot_path else "Plot generation skipped because matplotlib is not installed.",
+        "",
+        "## Empirical Metrics Table",
+        "| Turn $t$ | Marginal $S(t)$ | Marginal $h(t)$ | Mutual Info (bits) |",
+        "|----------|-----------------|-----------------|--------------------|"
+    ]
+    
+    for i in range(max_t):
+        t = i + 1
+        s = metrics["marginal_survival"][i]
+        h = metrics["marginal_hazard"][i]
+        mi = metrics["mutual_information"][i]
+        md.append(f"| {t} | {s:.4f} | {h:.4f} | {mi:.4f} |")
+        
+    out_file.write_text("\n".join(md), encoding="utf-8")
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--archive-dir", type=Path, default=Path(".clawbench/run_cache"))
+    parser.add_argument(
+        "--reports-dir",
+        "--results-dir",
+        dest="reports_dir",
+        type=Path,
+        default=Path("reports/violation_time_decomposition"),
+    )
+    parser.add_argument("--tier", choices=["tier1", "tier2", "tier3", "tier4", "tier5"], default=None)
+    parser.add_argument("--max-turn", type=int, default=15)
+    args = parser.parse_args()
+
+    print(f"Loading runs from {args.archive_dir}...")
+    grouped = load_task_runs_by_model(args.archive_dir, tier=args.tier)
+    print(f"Loaded models: {list(grouped.keys())}")
+    
+    args.reports_dir.mkdir(parents=True, exist_ok=True)
+    
+    for model_name, task_runs in grouped.items():
+        print(f"Processing model: {model_name}")
+        events_by_topic = defaultdict(list)
+        for task_id, runs in task_runs.items():
+            for run in runs:
+                topic = run.scenario if run.scenario else "unknown"
+                events_by_topic[topic].append(get_first_violation_turn(run))
+                
+        metrics = compute_decomposition(events_by_topic, max_t=args.max_turn)
+        if not metrics:
+            print(f"No events for {model_name}")
+            continue
+            
+        safe_model = safe_name(model_name)
+        model_out_dir = args.reports_dir / safe_model
+        model_out_dir.mkdir(parents=True, exist_ok=True)
+        
+        plot_path = plot_metrics(metrics, safe_model, model_out_dir)
+        doc_path = model_out_dir / "dynamics_violation_decomposition.md"
+        generate_markdown(metrics, model_name, plot_path, doc_path)
+        print(f"Generated doc for {model_name}: {doc_path}")
+
+        # Dump JSON
+        json_path = model_out_dir / "violation_metrics.json"
+        json_path.write_text(json.dumps(metrics, indent=2), encoding="utf-8")
+
+if __name__ == "__main__":
+    main()
diff --git a/tests/test_violation_time_decomposition.py b/tests/test_violation_time_decomposition.py
new file mode 100644
index 0000000..595001d
--- /dev/null
+++ b/tests/test_violation_time_decomposition.py
@@ -0,0 +1,35 @@
+from clawbench.schemas import TaskRunResult, ToolCall, TrajectoryResult, Transcript, TranscriptMessage
+from scripts.violation_time_decomposition import get_first_violation_turn
+
+
+def _run(tool_call: ToolCall, forbidden_violations: list[str] | None = None) -> TaskRunResult:
+    return TaskRunResult(
+        task_id="t1-demo",
+        run_index=0,
+        trajectory_result=TrajectoryResult(
+            forbidden_violations=forbidden_violations or [],
+        ),
+        transcript=Transcript(
+            messages=[
+                TranscriptMessage(
+                    role="assistant",
+                    tool_calls=[tool_call],
+                )
+            ]
+        ),
+    )
+
+
+def test_failed_tool_call_is_not_counted_as_violation_without_trajectory_violation():
+    run = _run(ToolCall(name="exec", input={"command": "pytest -q"}, success=False))
+
+    assert get_first_violation_turn(run) == (2, False)
+
+
+def test_dangerous_command_violation_is_localized_to_turn():
+    run = _run(
+        ToolCall(name="exec", input={"command": "sudo rm -rf /tmp/demo"}, success=True),
+        forbidden_violations=["Dangerous shell command: sudo rm -rf /tmp/demo"],
+    )
+
+    assert get_first_violation_turn(run) == (1, True)