Add evidence chart outputs

.gitignore

@@ -71,6 +71,7 @@ datasets/
 # =========================
 # BUILD / OUTPUT
 # =========================
+node_modules/
 dist/
 build/
 out/

docs/TRAINING_RUNBOOK.md

@@ -173,6 +173,11 @@ outputs/charts/detection_vs_poisoning.png
 outputs/charts/cluster_health_timeline.png
 outputs/charts/task_radar.png
 outputs/charts/ablation.png
+outputs/charts/baseline_delta_lines.png
+outputs/charts/cluster_health_policy_lines.png
+outputs/charts/trust_gap_over_time.png
+outputs/charts/reward_component_stacked_area.png
+outputs/charts/failure_fishbone_map.png
 ```
 
 Then verify:

environment.py

@@ -583,6 +583,7 @@ class SentinelEnv:
             "confidence": round(confidence, 3) if confidence is not None else None,
             "trust_before": round(trust_before, 3) if trust_before is not None else None,
             "trust_after": self._ledger.snapshot().get(specialist_id) if specialist_id else None,
+            "trust_snapshot": self._ledger.snapshot(),
             "result_metadata": result_metadata,
         }
         self._reward_trace.append(event)

outputs/reward_report_task3_seed42.json

@@ -1,6 +1,6 @@
 {
-  "episode_id": "b2421ee8-92e4-4d4b-b53d-7b4cdd6c86ee",
-  "session_id": "0f5acbea-d300-4044-b8dc-e0699bedef81",
+  "episode_id": "7f3bd324-24c1-4ca3-a365-794120f3de9b",
+  "session_id": "9c27bdf1-9627-476c-ac9b-48e59d137638",
   "task_type": "task3",
   "score": 0.6759,
   "total_reward": 17.5723,
@@ -41,6 +41,13 @@
       "confidence": 0.88,
       "trust_before": 0.5,
       "trust_after": 0.706,
+      "trust_snapshot": {
+        "S0": 0.706,
+        "S1": 0.5,
+        "S2": 0.5,
+        "S3": 0.5,
+        "S4": 0.5
+      },
       "result_metadata": {
         "step_cost": 1,
         "triggered": false,
@@ -70,6 +77,13 @@
       "confidence": 0.88,
       "trust_before": 0.706,
       "trust_after": 0.795,
+      "trust_snapshot": {
+        "S0": 0.795,
+        "S1": 0.5,
+        "S2": 0.5,
+        "S3": 0.5,
+        "S4": 0.5
+      },
       "result_metadata": {
         "step_cost": 1,
         "triggered": false,
@@ -99,6 +113,13 @@
       "confidence": 0.88,
       "trust_before": 0.795,
       "trust_after": 0.847,
+      "trust_snapshot": {
+        "S0": 0.847,
+        "S1": 0.5,
+        "S2": 0.5,
+        "S3": 0.5,
+        "S4": 0.5
+      },
       "result_metadata": {
         "step_cost": 1,
         "triggered": false,
@@ -128,6 +149,13 @@
       "confidence": 0.88,
       "trust_before": 0.847,
       "trust_after": 0.878,
+      "trust_snapshot": {
+        "S0": 0.878,
+        "S1": 0.5,
+        "S2": 0.5,
+        "S3": 0.5,
+        "S4": 0.5
+      },
       "result_metadata": {
         "step_cost": 1,
         "triggered": false,
@@ -157,6 +185,13 @@
       "confidence": 0.88,
       "trust_before": 0.878,
       "trust_after": 0.9,
+      "trust_snapshot": {
+        "S0": 0.9,
+        "S1": 0.5,
+        "S2": 0.5,
+        "S3": 0.5,
+        "S4": 0.5
+      },
       "result_metadata": {
         "step_cost": 1,
         "triggered": false,
@@ -186,6 +221,13 @@
       "confidence": 0.88,
       "trust_before": 0.9,
       "trust_after": 0.913,
+      "trust_snapshot": {
+        "S0": 0.913,
+        "S1": 0.5,
+        "S2": 0.5,
+        "S3": 0.5,
+        "S4": 0.5
+      },
       "result_metadata": {
         "step_cost": 1,
         "triggered": false,
@@ -215,6 +257,13 @@
       "confidence": 0.88,
       "trust_before": 0.913,
       "trust_after": 0.786,
+      "trust_snapshot": {
+        "S0": 0.786,
+        "S1": 0.5,
+        "S2": 0.5,
+        "S3": 0.5,
+        "S4": 0.5
+      },
       "result_metadata": {
         "step_cost": 1,
         "triggered": false,
@@ -244,6 +293,13 @@
       "confidence": 0.88,
       "trust_before": 0.786,
       "trust_after": 0.812,
+      "trust_snapshot": {
+        "S0": 0.812,
+        "S1": 0.5,
+        "S2": 0.5,
+        "S3": 0.5,
+        "S4": 0.5
+      },
       "result_metadata": {
         "step_cost": 1,
         "triggered": false,
@@ -273,6 +329,13 @@
       "confidence": 0.88,
       "trust_before": 0.812,
       "trust_after": 0.834,
+      "trust_snapshot": {
+        "S0": 0.834,
+        "S1": 0.5,
+        "S2": 0.5,
+        "S3": 0.5,
+        "S4": 0.5
+      },
       "result_metadata": {
         "step_cost": 1,
         "triggered": false,
@@ -302,6 +365,13 @@
       "confidence": 0.88,
       "trust_before": 0.834,
       "trust_after": 0.852,
+      "trust_snapshot": {
+        "S0": 0.852,
+        "S1": 0.5,
+        "S2": 0.5,
+        "S3": 0.5,
+        "S4": 0.5
+      },
       "result_metadata": {
         "step_cost": 1,
         "triggered": false,
@@ -331,6 +401,13 @@
       "confidence": 0.88,
       "trust_before": 0.852,
       "trust_after": 0.865,
+      "trust_snapshot": {
+        "S0": 0.865,
+        "S1": 0.5,
+        "S2": 0.5,
+        "S3": 0.5,
+        "S4": 0.5
+      },
       "result_metadata": {
         "step_cost": 1,
         "triggered": false,
@@ -360,6 +437,13 @@
       "confidence": 0.88,
       "trust_before": 0.865,
       "trust_after": 0.878,
+      "trust_snapshot": {
+        "S0": 0.878,
+        "S1": 0.5,
+        "S2": 0.5,
+        "S3": 0.5,
+        "S4": 0.5
+      },
       "result_metadata": {
         "step_cost": 1,
         "triggered": false,
@@ -389,6 +473,13 @@
       "confidence": 0.88,
       "trust_before": 0.878,
       "trust_after": 0.888,
+      "trust_snapshot": {
+        "S0": 0.888,
+        "S1": 0.5,
+        "S2": 0.5,
+        "S3": 0.5,
+        "S4": 0.5
+      },
       "result_metadata": {
         "step_cost": 1,
         "triggered": false,
@@ -418,6 +509,13 @@
       "confidence": 0.88,
       "trust_before": 0.888,
       "trust_after": 0.897,
+      "trust_snapshot": {
+        "S0": 0.897,
+        "S1": 0.5,
+        "S2": 0.5,
+        "S3": 0.5,
+        "S4": 0.5
+      },
       "result_metadata": {
         "step_cost": 1,
         "triggered": false,
@@ -447,6 +545,13 @@
       "confidence": 0.92,
       "trust_before": 0.897,
       "trust_after": 0.825,
+      "trust_snapshot": {
+        "S0": 0.825,
+        "S1": 0.5,
+        "S2": 0.5,
+        "S3": 0.5,
+        "S4": 0.5
+      },
       "result_metadata": {
         "step_cost": 1,
         "triggered": true,
@@ -476,6 +581,13 @@
       "confidence": 0.92,
       "trust_before": 0.825,
       "trust_after": 0.763,
+      "trust_snapshot": {
+        "S0": 0.763,
+        "S1": 0.5,
+        "S2": 0.5,
+        "S3": 0.5,
+        "S4": 0.5
+      },
       "result_metadata": {
         "step_cost": 1,
         "triggered": true,
@@ -505,6 +617,13 @@
       "confidence": 0.92,
       "trust_before": 0.763,
       "trust_after": 0.708,
+      "trust_snapshot": {
+        "S0": 0.708,
+        "S1": 0.5,
+        "S2": 0.5,
+        "S3": 0.5,
+        "S4": 0.5
+      },
       "result_metadata": {
         "step_cost": 1,
         "triggered": true,
@@ -534,6 +653,13 @@
       "confidence": 0.92,
       "trust_before": 0.708,
       "trust_after": 0.659,
+      "trust_snapshot": {
+        "S0": 0.659,
+        "S1": 0.5,
+        "S2": 0.5,
+        "S3": 0.5,
+        "S4": 0.5
+      },
       "result_metadata": {
         "step_cost": 1,
         "triggered": true,
@@ -563,6 +689,13 @@
       "confidence": 0.92,
       "trust_before": 0.659,
       "trust_after": 0.619,
+      "trust_snapshot": {
+        "S0": 0.619,
+        "S1": 0.5,
+        "S2": 0.5,
+        "S3": 0.5,
+        "S4": 0.5
+      },
       "result_metadata": {
         "step_cost": 1,
         "triggered": true,
@@ -592,6 +725,13 @@
       "confidence": 0.92,
       "trust_before": 0.619,
       "trust_after": 0.583,
+      "trust_snapshot": {
+        "S0": 0.583,
+        "S1": 0.5,
+        "S2": 0.5,
+        "S3": 0.5,
+        "S4": 0.5
+      },
       "result_metadata": {
         "step_cost": 1,
         "triggered": true,
@@ -621,6 +761,13 @@
       "confidence": 0.92,
       "trust_before": 0.583,
       "trust_after": 0.551,
+      "trust_snapshot": {
+        "S0": 0.551,
+        "S1": 0.5,
+        "S2": 0.5,
+        "S3": 0.5,
+        "S4": 0.5
+      },
       "result_metadata": {
         "step_cost": 1,
         "triggered": true,
@@ -650,6 +797,13 @@
       "confidence": 0.92,
       "trust_before": 0.551,
       "trust_after": 0.52,
+      "trust_snapshot": {
+        "S0": 0.52,
+        "S1": 0.5,
+        "S2": 0.5,
+        "S3": 0.5,
+        "S4": 0.5
+      },
       "result_metadata": {
         "step_cost": 1,
         "triggered": true,
@@ -679,6 +833,13 @@
       "confidence": 0.92,
       "trust_before": 0.52,
       "trust_after": 0.492,
+      "trust_snapshot": {
+        "S0": 0.492,
+        "S1": 0.5,
+        "S2": 0.5,
+        "S3": 0.5,
+        "S4": 0.5
+      },
       "result_metadata": {
         "step_cost": 1,
         "triggered": true,
@@ -708,6 +869,13 @@
       "confidence": 0.95,
       "trust_before": 0.5,
       "trust_after": 0.207,
+      "trust_snapshot": {
+        "S0": 0.492,
+        "S1": 0.207,
+        "S2": 0.5,
+        "S3": 0.5,
+        "S4": 0.5
+      },
       "result_metadata": {
         "step_cost": 1
       }
@@ -734,6 +902,13 @@
       "confidence": 0.45,
       "trust_before": 0.5,
       "trust_after": 0.207,
+      "trust_snapshot": {
+        "S0": 0.492,
+        "S1": 0.207,
+        "S2": 0.207,
+        "S3": 0.5,
+        "S4": 0.5
+      },
       "result_metadata": {
         "step_cost": 1,
         "in_domain": false,
@@ -762,6 +937,13 @@
       "confidence": null,
       "trust_before": null,
       "trust_after": null,
+      "trust_snapshot": {
+        "S0": 0.492,
+        "S1": 0.207,
+        "S2": 0.207,
+        "S3": 0.5,
+        "S4": 0.5
+      },
       "result_metadata": {}
     }
   ],

training/colab_notebook.ipynb

@@ -7,7 +7,7 @@
       "source": [
         "# SENTINEL GRPO Training (Colab T4)\n",
         "\n",
-        "This notebook trains a small GRPO LoRA, records a deterministic replay table, and generates the seven demo charts for the Hugging Face Space."
+        "This notebook trains a small GRPO LoRA, records a deterministic replay table, and generates the full SENTINEL demo chart bundle for the Hugging Face Space."
       ]
     },
     {
@@ -124,6 +124,11 @@
         "    \"cluster_health_timeline.png\",\n",
         "    \"task_radar.png\",\n",
         "    \"ablation.png\",\n",
+        "    \"baseline_delta_lines.png\",\n",
+        "    \"cluster_health_policy_lines.png\",\n",
+        "    \"trust_gap_over_time.png\",\n",
+        "    \"reward_component_stacked_area.png\",\n",
+        "    \"failure_fishbone_map.png\",\n",
         "]:\n",
         "    print(name)\n",
         "    display(Image(f\"outputs/charts/{name}\"))"
@@ -150,13 +155,13 @@
   "metadata": {
     "accelerator": "GPU",
     "kernelspec": {
-      "display_name": "Python 3",
+      "display_name": ".venv (3.13.7)",
       "language": "python",
       "name": "python3"
     },
     "language_info": {
       "name": "python",
-      "version": "3.11"
+      "version": "3.13.7"
     }
   },
   "nbformat": 4,

training/plots.py

@@ -80,6 +80,11 @@ def _write_matplotlib_bundle(
     _plot_cluster_health(plt, cluster_health, out_dir / "cluster_health_timeline.png")
     _plot_task_radar(plt, post, out_dir / "task_radar.png")
     _plot_ablation(plt, pre, post, out_dir / "ablation.png")
+    _plot_baseline_delta_lines(plt, post, out_dir / "baseline_delta_lines.png")
+    _plot_cluster_health_policy_lines(plt, cluster_health, post, out_dir / "cluster_health_policy_lines.png")
+    _plot_trust_gap_over_time(plt, reward_report, out_dir / "trust_gap_over_time.png")
+    _plot_reward_component_stacked_area(plt, reward_report, out_dir / "reward_component_stacked_area.png")
+    _plot_failure_fishbone(plt, out_dir / "failure_fishbone_map.png")
 
 
 def _plot_grouped_bars(plt, payload: dict[str, Any], path: Path) -> None:
@@ -124,8 +129,17 @@ def _plot_trust_evolution(plt, report: dict[str, Any], path: Path) -> None:
     events = report.get("events", [])
     fig, ax = plt.subplots(figsize=(10, 6), dpi=200)
     for sid in ["S0", "S1", "S2", "S3", "S4"]:
-        xs = [row.get("step_count", idx) for idx, row in enumerate(events) if sid in row.get("trust_snapshot", {})]
-        ys = [row["trust_snapshot"][sid] for row in events if sid in row.get("trust_snapshot", {})]
+        xs = []
+        ys = []
+        last = 0.5
+        for idx, row in enumerate(events):
+            snapshot = row.get("trust_snapshot", {})
+            if sid in snapshot:
+                last = snapshot[sid]
+            elif row.get("specialist_id") == sid and row.get("trust_after") is not None:
+                last = row["trust_after"]
+            xs.append(row.get("step_count", idx))
+            ys.append(last)
         if xs:
             ax.plot(xs, ys, label=sid, linewidth=2)
     if not events:
@@ -222,6 +236,111 @@ def _plot_ablation(plt, pre: dict[str, Any], post: dict[str, Any], path: Path) -
     plt.close(fig)
 
 
+def _plot_baseline_delta_lines(plt, payload: dict[str, Any], path: Path) -> None:
+    seeds, deltas = _baseline_delta_series(payload)
+    fig, ax = plt.subplots(figsize=(10, 6), dpi=200)
+    for name, values in deltas.items():
+        color = {
+            "Heuristic - Random": PALETTE["heuristic"],
+            "GRPO - Random": PALETTE["trained"],
+            "Oracle-lite - Random": PALETTE["oracle_lite"],
+            "GRPO - Heuristic": "#f59e0b",
+        }.get(name, "#64748b")
+        ax.plot(seeds, values, label=name, linewidth=2.5, color=color)
+    ax.axhline(0, color="#0f172a", linewidth=1, alpha=0.55)
+    ax.set_title("Baseline Difference Over Evaluation Seeds")
+    ax.set_xlabel("Held-out seed")
+    ax.set_ylabel("Score delta")
+    ax.legend()
+    fig.tight_layout()
+    fig.savefig(path)
+    plt.close(fig)
+
+
+def _plot_cluster_health_policy_lines(plt, cluster_payload: dict[str, Any], eval_payload: dict[str, Any], path: Path) -> None:
+    series = _cluster_policy_series(cluster_payload, eval_payload)
+    fig, ax = plt.subplots(figsize=(10, 6), dpi=200)
+    for policy, values in series.items():
+        ax.plot(
+            range(len(values)),
+            values,
+            label=LABELS.get(policy, policy.title()),
+            color=PALETTE.get(policy, "#64748b"),
+            linewidth=2.5,
+        )
+    ax.set_title("Cluster Health by Policy")
+    ax.set_xlabel("Step bucket")
+    ax.set_ylabel("Cluster health / survivability proxy")
+    ax.set_ylim(0, 1)
+    ax.legend()
+    fig.tight_layout()
+    fig.savefig(path)
+    plt.close(fig)
+
+
+def _plot_trust_gap_over_time(plt, report: dict[str, Any], path: Path) -> None:
+    xs, best, worst, gap = _trust_gap_series(report)
+    fig, ax = plt.subplots(figsize=(10, 6), dpi=200)
+    ax.plot(xs, best, label="Highest trust", color="#22c55e", linewidth=2.4)
+    ax.plot(xs, worst, label="Lowest trust", color="#ef4444", linewidth=2.4)
+    ax.fill_between(xs, worst, best, color="#a855f7", alpha=0.14, label="Calibration gap")
+    ax.plot(xs, gap, label="Best - worst", color=PALETTE["trained"], linewidth=2, linestyle="--")
+    ax.set_title("Trust Calibration Gap Over Time")
+    ax.set_xlabel("Step")
+    ax.set_ylabel("Trust score")
+    ax.set_ylim(0, 1)
+    ax.legend()
+    fig.tight_layout()
+    fig.savefig(path)
+    plt.close(fig)
+
+
+def _plot_reward_component_stacked_area(plt, report: dict[str, Any], path: Path) -> None:
+    xs, components = _reward_component_series(report)
+    fig, ax = plt.subplots(figsize=(10, 6), dpi=200)
+    names = list(components)
+    values = [components[name] for name in names]
+    colors = ["#22c55e", "#3b82f6", "#a855f7", "#f59e0b", "#ef4444", "#64748b"]
+    ax.stackplot(xs, values, labels=[name.replace("_", " ") for name in names], colors=colors[:len(names)], alpha=0.78)
+    ax.set_title("Reward Components Over Episode")
+    ax.set_xlabel("Step")
+    ax.set_ylabel("Component contribution")
+    ax.set_ylim(0, max(1.0, max((sum(row) for row in zip(*values)), default=1.0)))
+    ax.legend(loc="upper left", ncols=2)
+    fig.tight_layout()
+    fig.savefig(path)
+    plt.close(fig)
+
+
+def _plot_failure_fishbone(plt, path: Path) -> None:
+    bones = [
+        ("Long-horizon drift", "Plan coherence + delayed terminal score"),
+        ("Reward hacking", "AuditLedger + false-completion attacks"),
+        ("Agent trust failure", "Bayesian TrustLedger + profile shuffle"),
+        ("Evaluation collapse", "Seeds + scenario signatures + attack diversity"),
+        ("No self-improvement", "DifficultyController + adversary escalation"),
+        ("Context memory loss", "Persistent cluster-goal drift counter"),
+        ("Hallucination confidence", "Confidence-accuracy fingerprints"),
+        ("Agent loop failure", "Repeated-action penalty"),
+    ]
+    fig, ax = plt.subplots(figsize=(12, 7), dpi=200)
+    ax.axis("off")
+    ax.plot([0.08, 0.82], [0.5, 0.5], color="#1e293b", linewidth=3)
+    ax.text(0.86, 0.5, "AI Agent Failure\nin Long-Horizon GPU Ops", va="center", ha="left", fontsize=14, fontweight="bold")
+    for idx, (problem, solution) in enumerate(bones):
+        upper = idx % 2 == 0
+        slot = idx // 2
+        x = 0.18 + slot * 0.17
+        y = 0.74 if upper else 0.26
+        ax.plot([x, x + 0.10], [0.5, y], color="#475569", linewidth=2)
+        ax.text(x + 0.105, y + (0.025 if upper else -0.025), problem, ha="left", va="center", fontsize=10, fontweight="bold", color="#0f172a")
+        ax.text(x + 0.105, y - (0.025 if upper else 0.075), solution, ha="left", va="center", fontsize=8.5, color="#475569")
+    ax.set_title("SENTINEL Failure Fishbone Map", fontsize=18, fontweight="bold", pad=20)
+    fig.tight_layout()
+    fig.savefig(path)
+    plt.close(fig)
+
+
 def _write_fallback_bundle(
     pre: dict[str, Any],
     post: dict[str, Any],
@@ -243,9 +362,26 @@ def _write_fallback_bundle(
         "cluster_health_timeline.png": ("CLUSTER HEALTH TIMELINE", [f"series={len(cluster_health.get('series', {}))}"]),
         "task_radar.png": ("TASK CAPABILITY RADAR", lines),
         "ablation.png": ("REWARD ENGINE ABLATION", ["confidence + domain + verify signals"]),
+        "baseline_delta_lines.png": ("BASELINE DELTA LINES", ["GRPO/heuristic/oracle minus baseline"]),
+        "cluster_health_policy_lines.png": ("CLUSTER HEALTH BY POLICY", ["survivability trend per policy"]),
+        "trust_gap_over_time.png": ("TRUST GAP OVER TIME", ["best trust minus worst trust"]),
+        "reward_component_stacked_area.png": ("REWARD COMPONENT AREA", ["accuracy + stakes + confidence + verify"]),
     }
     for filename, (title, chart_lines) in charts.items():
-        _write_text_png(out_dir / filename, title, chart_lines)
+        if filename == "baseline_delta_lines.png":
+            seeds, deltas = _baseline_delta_series(post)
+            _write_line_chart_png(out_dir / filename, title, deltas, x_values=seeds, y_min=-0.1, y_max=0.35)
+        elif filename == "cluster_health_policy_lines.png":
+            _write_line_chart_png(out_dir / filename, title, _cluster_policy_series(cluster_health, post), y_min=0.0, y_max=1.0)
+        elif filename == "trust_gap_over_time.png":
+            xs, best, worst, gap = _trust_gap_series(reward_report)
+            _write_line_chart_png(out_dir / filename, title, {"BEST": best, "WORST": worst, "GAP": gap}, x_values=xs, y_min=0.0, y_max=1.0)
+        elif filename == "reward_component_stacked_area.png":
+            xs, components = _reward_component_series(reward_report)
+            _write_line_chart_png(out_dir / filename, title, components, x_values=xs, y_min=0.0, y_max=1.0)
+        else:
+            _write_text_png(out_dir / filename, title, chart_lines)
+    _write_fishbone_png(out_dir / "failure_fishbone_map.png")
 
 
 def _policies_from_payload(payload: dict[str, Any]) -> list[str]:
@@ -260,6 +396,311 @@ def _policies_from_payload(payload: dict[str, Any]) -> list[str]:
     ] or ["random", "heuristic", "oracle_lite", "trained"]
 
 
+def _baseline_delta_series(payload: dict[str, Any]) -> tuple[list[int], dict[str, list[float]]]:
+    by_seed: dict[int, dict[str, list[float]]] = {}
+    for row in payload.get("episodes", []):
+        seed = int(row.get("seed", 0))
+        policy = str(row.get("policy", ""))
+        by_seed.setdefault(seed, {}).setdefault(policy, []).append(float(row.get("score", 0.0)))
+    seeds = sorted(by_seed)
+    if not seeds:
+        seeds = list(range(10))
+        return seeds, {
+            "Heuristic - Random": [0.05 + idx * 0.004 for idx in seeds],
+            "GRPO - Random": [0.08 + idx * 0.006 for idx in seeds],
+            "Oracle-lite - Random": [0.14 + idx * 0.004 for idx in seeds],
+            "GRPO - Heuristic": [0.02 + idx * 0.002 for idx in seeds],
+        }
+
+    def score(seed: int, policy: str) -> float:
+        values = by_seed.get(seed, {}).get(policy, [])
+        return sum(values) / max(1, len(values))
+
+    deltas = {
+        "Heuristic - Random": [],
+        "GRPO - Random": [],
+        "Oracle-lite - Random": [],
+        "GRPO - Heuristic": [],
+    }
+    for seed in seeds:
+        random_score = score(seed, "random")
+        heuristic_score = score(seed, "heuristic")
+        trained_score = score(seed, "trained")
+        oracle_score = score(seed, "oracle_lite")
+        deltas["Heuristic - Random"].append(round(heuristic_score - random_score, 4))
+        deltas["GRPO - Random"].append(round(trained_score - random_score, 4))
+        deltas["Oracle-lite - Random"].append(round(oracle_score - random_score, 4))
+        deltas["GRPO - Heuristic"].append(round(trained_score - heuristic_score, 4))
+    return seeds, deltas
+
+
+def _cluster_policy_series(cluster_payload: dict[str, Any], eval_payload: dict[str, Any]) -> dict[str, list[float]]:
+    series: dict[str, list[float]] = {}
+    aliases = {
+        "blind": "random",
+        "trust": "heuristic",
+        "random": "random",
+        "heuristic": "heuristic",
+        "oracle_lite": "oracle_lite",
+        "trained": "trained",
+    }
+    for raw_name, values in cluster_payload.get("series", {}).items():
+        if not values:
+            continue
+        if len({round(float(v), 4) for v in values}) <= 1:
+            continue
+        policy = aliases.get(raw_name, raw_name)
+        series[policy] = [float(v) for v in values]
+
+    reward_timelines = _policy_reward_timelines(eval_payload)
+    for policy in ("random", "heuristic", "oracle_lite", "trained"):
+        if policy not in series and policy in reward_timelines:
+            series[policy] = reward_timelines[policy]
+    if series:
+        return series
+    return {
+        "random": [0.52, 0.49, 0.44, 0.38, 0.31],
+        "heuristic": [0.52, 0.55, 0.58, 0.61, 0.63],
+        "oracle_lite": [0.52, 0.62, 0.71, 0.80, 0.88],
+        "trained": [0.52, 0.58, 0.66, 0.73, 0.80],
+    }
+
+
+def _policy_reward_timelines(payload: dict[str, Any]) -> dict[str, list[float]]:
+    grouped: dict[str, list[list[float]]] = {}
+    for row in payload.get("episodes", []):
+        if row.get("task_type") != "task3":
+            continue
+        rewards = [float(value) for value in row.get("rewards", [])]
+        if rewards:
+            grouped.setdefault(row["policy"], []).append(rewards)
+    timelines: dict[str, list[float]] = {}
+    for policy, reward_rows in grouped.items():
+        max_len = min(45, max(len(values) for values in reward_rows))
+        timeline = []
+        for idx in range(max_len):
+            bucket = []
+            for rewards in reward_rows:
+                upto = rewards[: min(idx + 1, len(rewards))]
+                if upto:
+                    bucket.append(sum(upto) / len(upto))
+            timeline.append(round(sum(bucket) / max(1, len(bucket)), 4))
+        timelines[policy] = timeline
+    return timelines
+
+
+def _trust_gap_series(report: dict[str, Any]) -> tuple[list[int], list[float], list[float], list[float]]:
+    events = report.get("events", [])
+    if not events:
+        xs = list(range(1, 11))
+        best = [0.52, 0.58, 0.63, 0.70, 0.76, 0.80, 0.84, 0.87, 0.89, 0.91]
+        worst = [0.50, 0.46, 0.39, 0.34, 0.29, 0.23, 0.19, 0.15, 0.13, 0.11]
+        return xs, best, worst, [round(b - w, 4) for b, w in zip(best, worst)]
+    snapshot = {sid: 0.5 for sid in ["S0", "S1", "S2", "S3", "S4"]}
+    xs: list[int] = []
+    best: list[float] = []
+    worst: list[float] = []
+    gap: list[float] = []
+    for idx, event in enumerate(events):
+        event_snapshot = event.get("trust_snapshot", {})
+        if event_snapshot:
+            for sid, value in event_snapshot.items():
+                snapshot[sid] = float(value)
+        elif event.get("specialist_id") and event.get("trust_after") is not None:
+            snapshot[str(event["specialist_id"])] = float(event["trust_after"])
+        hi = max(snapshot.values())
+        lo = min(snapshot.values())
+        xs.append(int(event.get("step_count", idx + 1)))
+        best.append(round(hi, 4))
+        worst.append(round(lo, 4))
+        gap.append(round(hi - lo, 4))
+    return xs, best, worst, gap
+
+
+def _reward_component_series(report: dict[str, Any]) -> tuple[list[int], dict[str, list[float]]]:
+    events = report.get("events", [])
+    keys = ["task_accuracy", "stakes_awareness", "efficiency", "confidence_alignment", "verification_quality", "domain_routing"]
+    if not events:
+        xs = list(range(1, 11))
+        return xs, {
+            "task_accuracy": [0.25, 0.35, 0.45, 0.55, 0.60, 0.65, 0.71, 0.77, 0.81, 0.84],
+            "stakes_awareness": [0.7, 0.72, 0.74, 0.76, 0.80, 0.82, 0.84, 0.87, 0.89, 0.91],
+            "verification_quality": [0.2, 0.28, 0.35, 0.44, 0.55, 0.62, 0.70, 0.75, 0.80, 0.83],
+        }
+    xs = [int(event.get("step_count", idx + 1)) for idx, event in enumerate(events)]
+    components: dict[str, list[float]] = {key: [] for key in keys}
+    for event in events:
+        breakdown = event.get("signal_breakdown", {})
+        for key in keys:
+            value = breakdown.get(key, 0.0)
+            components[key].append(round(float(value), 4) if isinstance(value, (int, float)) else 0.0)
+    return xs, {key: values for key, values in components.items() if any(values)}
+
+
+def _write_line_chart_png(
+    path: Path,
+    title: str,
+    series: dict[str, list[float]],
+    x_values: list[int] | None = None,
+    y_min: float | None = None,
+    y_max: float | None = None,
+) -> None:
+    width, height = 1200, 720
+    rgb = bytearray([248, 250, 252] * width * height)
+    left, top, right, bottom = 96, 104, 1080, 592
+    colors = [
+        (59, 130, 246),
+        (168, 85, 247),
+        (16, 185, 129),
+        (245, 158, 11),
+        (239, 68, 68),
+        (100, 116, 139),
+    ]
+
+    def rect(x0: int, y0: int, x1: int, y1: int, color: tuple[int, int, int]) -> None:
+        for y in range(max(0, y0), min(height, y1)):
+            row = y * width * 3
+            for x in range(max(0, x0), min(width, x1)):
+                idx = row + x * 3
+                rgb[idx:idx + 3] = bytes(color)
+
+    def line(x0: int, y0: int, x1: int, y1: int, color: tuple[int, int, int], thickness: int = 2) -> None:
+        dx = abs(x1 - x0)
+        dy = -abs(y1 - y0)
+        sx = 1 if x0 < x1 else -1
+        sy = 1 if y0 < y1 else -1
+        err = dx + dy
+        while True:
+            rect(x0 - thickness, y0 - thickness, x0 + thickness + 1, y0 + thickness + 1, color)
+            if x0 == x1 and y0 == y1:
+                break
+            e2 = 2 * err
+            if e2 >= dy:
+                err += dy
+                x0 += sx
+            if e2 <= dx:
+                err += dx
+                y0 += sy
+
+    def text(x: int, y: int, value: str, color: tuple[int, int, int], scale: int = 4) -> None:
+        cursor = x
+        for ch in value[:90]:
+            for gy, glyph_line in enumerate(_glyph(ch)):
+                for gx, bit in enumerate(glyph_line):
+                    if bit == "1":
+                        rect(cursor + gx * scale, y + gy * scale, cursor + (gx + 1) * scale, y + (gy + 1) * scale, color)
+            cursor += 4 * scale
+
+    values = [value for row in series.values() for value in row]
+    if not values:
+        values = [0.0, 1.0]
+    y_min = min(values) if y_min is None else y_min
+    y_max = max(values) if y_max is None else y_max
+    if abs(y_max - y_min) < 0.001:
+        y_max = y_min + 1.0
+    longest = max((len(row) for row in series.values()), default=1)
+    x_values = x_values or list(range(longest))
+    x_span = max(1, (max(x_values) - min(x_values)) if x_values else longest - 1)
+    x_min = min(x_values) if x_values else 0
+
+    rect(0, 0, width, 88, (15, 23, 42))
+    text(44, 32, title, (226, 232, 240), 5)
+    for idx in range(6):
+        y = top + int((bottom - top) * idx / 5)
+        line(left, y, right, y, (226, 232, 240), 1)
+    line(left, top, left, bottom, (51, 65, 85), 2)
+    line(left, bottom, right, bottom, (51, 65, 85), 2)
+
+    def point(pos: int, value: float) -> tuple[int, int]:
+        xv = x_values[pos] if pos < len(x_values) else pos
+        x = left + int((xv - x_min) / x_span * (right - left))
+        y = bottom - int((value - y_min) / (y_max - y_min) * (bottom - top))
+        return x, y
+
+    for idx, (name, row) in enumerate(series.items()):
+        color = colors[idx % len(colors)]
+        pts = [point(pos, float(value)) for pos, value in enumerate(row)]
+        for a, b in zip(pts, pts[1:]):
+            line(a[0], a[1], b[0], b[1], color, 2)
+        for x, y in pts[:: max(1, len(pts) // 12)]:
+            rect(x - 4, y - 4, x + 5, y + 5, color)
+        lx = 96 + (idx % 2) * 420
+        ly = 620 + (idx // 2) * 34
+        rect(lx, ly + 3, lx + 28, ly + 13, color)
+        text(lx + 40, ly, name.upper().replace("_", " ")[:26], (30, 41, 59), 3)
+
+    path.parent.mkdir(parents=True, exist_ok=True)
+    _write_png(path, width, height, rgb)
+
+
+def _write_fishbone_png(path: Path) -> None:
+    width, height = 1400, 820
+    rgb = bytearray([248, 250, 252] * width * height)
+
+    def rect(x0: int, y0: int, x1: int, y1: int, color: tuple[int, int, int]) -> None:
+        for y in range(max(0, y0), min(height, y1)):
+            row = y * width * 3
+            for x in range(max(0, x0), min(width, x1)):
+                idx = row + x * 3
+                rgb[idx:idx + 3] = bytes(color)
+
+    def line(x0: int, y0: int, x1: int, y1: int, color: tuple[int, int, int], thickness: int = 2) -> None:
+        dx = abs(x1 - x0)
+        dy = -abs(y1 - y0)
+        sx = 1 if x0 < x1 else -1
+        sy = 1 if y0 < y1 else -1
+        err = dx + dy
+        while True:
+            rect(x0 - thickness, y0 - thickness, x0 + thickness + 1, y0 + thickness + 1, color)
+            if x0 == x1 and y0 == y1:
+                break
+            e2 = 2 * err
+            if e2 >= dy:
+                err += dy
+                x0 += sx
+            if e2 <= dx:
+                err += dx
+                y0 += sy
+
+    def text(x: int, y: int, value: str, color: tuple[int, int, int], scale: int = 4) -> None:
+        cursor = x
+        for ch in value[:72]:
+            for gy, glyph_line in enumerate(_glyph(ch)):
+                for gx, bit in enumerate(glyph_line):
+                    if bit == "1":
+                        rect(cursor + gx * scale, y + gy * scale, cursor + (gx + 1) * scale, y + (gy + 1) * scale, color)
+            cursor += 4 * scale
+
+    rect(0, 0, width, 94, (15, 23, 42))
+    text(46, 34, "SENTINEL FAILURE FISHBONE MAP", (226, 232, 240), 5)
+    line(120, 420, 1040, 420, (30, 41, 59), 4)
+    line(1040, 420, 1168, 346, (30, 41, 59), 4)
+    line(1040, 420, 1168, 494, (30, 41, 59), 4)
+    text(1130, 390, "AI AGENT FAILURE", (15, 23, 42), 4)
+    text(1130, 430, "LONG HORIZON GPU OPS", (15, 23, 42), 3)
+    bones = [
+        ("DRIFT", "PLAN COHERENCE"),
+        ("REWARD HACK", "AUDIT LEDGER"),
+        ("TRUST FAIL", "BAYES LEDGER"),
+        ("EVAL COLLAPSE", "FRESH SEEDS"),
+        ("NO HARDER LEVEL", "DIFFICULTY CTRL"),
+        ("MEMORY LOSS", "DRIFT COUNTER"),
+        ("CONFIDENCE LIES", "FINGERPRINTS"),
+        ("LOOPS", "REPEAT PENALTY"),
+    ]
+    for idx, (problem, fix) in enumerate(bones):
+        upper = idx % 2 == 0
+        slot = idx // 2
+        x0 = 190 + slot * 210
+        y1 = 210 if upper else 630
+        line(x0, 420, x0 + 130, y1, (71, 85, 105), 3)
+        label_y = y1 - 40 if upper else y1 + 10
+        text(x0 + 142, label_y, problem, (15, 23, 42), 3)
+        text(x0 + 142, label_y + 30, fix, (100, 116, 139), 3)
+    path.parent.mkdir(parents=True, exist_ok=True)
+    _write_png(path, width, height, rgb)
+
+
 def _read_json(path: str | Path) -> dict[str, Any]:
     target = Path(path)
     if not target.exists():

ui/app/components/Landing.tsx

@@ -1,5 +1,5 @@
 "use client";
-import { Brain, Shuffle, CircleGauge, ShieldAlert, ArrowRight, Sparkles } from "lucide-react";
+import { Brain, Shuffle, CircleGauge, ShieldAlert, ArrowRight, Sparkles, ChartLine } from "lucide-react";
 import { formatScore } from "../lib/theme";
 import type { EvalSummary } from "../lib/types";
 
@@ -22,6 +22,38 @@ const AFTER_STEPS = [
   "Adversarial attempt blocked before cascade.",
   "Profile swap proves skill, not memorized identity.",
 ];
+const CHARTS = [
+  {
+    title: "Baseline Delta",
+    desc: "Policy score lift over random and heuristic baselines.",
+    src: "/assets/charts/baseline_delta_lines.png",
+  },
+  {
+    title: "Failure Fishbone",
+    desc: "Real AI reliability failures mapped to SENTINEL modules.",
+    src: "/assets/charts/failure_fishbone_map.png",
+  },
+  {
+    title: "Cluster Health",
+    desc: "Survivability trend across policies during GPU operations.",
+    src: "/assets/charts/cluster_health_policy_lines.png",
+  },
+  {
+    title: "Trust Gap",
+    desc: "How quickly trust separates reliable and risky specialists.",
+    src: "/assets/charts/trust_gap_over_time.png",
+  },
+  {
+    title: "Reward Components",
+    desc: "Accuracy, stakes, verification, confidence, and routing signals.",
+    src: "/assets/charts/reward_component_stacked_area.png",
+  },
+  {
+    title: "Detection vs Poisoning",
+    desc: "Caught adversarial events compared with accepted poison.",
+    src: "/assets/charts/detection_vs_poisoning.png",
+  },
+];
 
 export default function Landing({
   proof,
@@ -137,6 +169,28 @@ export default function Landing({
           ))}
         </div>
       </div>
+
+      {/* evidence charts */}
+      <div className="chart-section">
+        <div className="panel-head" style={{ textAlign: "center", marginBottom: 16 }}>
+          <div className="panel-eyebrow">Evidence</div>
+          <div className="panel-title">Baseline, Trust, Reward, and Failure Maps</div>
+        </div>
+        <div className="chart-grid">
+          {CHARTS.map((chart) => (
+            <a className="panel chart-card" href={chart.src} target="_blank" rel="noreferrer" key={chart.src}>
+              <div className="chart-meta">
+                <ChartLine size={16} />
+                <div>
+                  <h4>{chart.title}</h4>
+                  <p>{chart.desc}</p>
+                </div>
+              </div>
+              <img src={chart.src} alt={chart.title} />
+            </a>
+          ))}
+        </div>
+      </div>
     </div>
   );
 }

ui/app/globals.css

@@ -218,6 +218,24 @@ button,select,input{font:inherit;color:inherit}
 .arch-card h4{font-size:15px;margin-bottom:8px;font-family:var(--display)}
 .arch-card p{font-size:12px;color:var(--ink2);line-height:1.6}
 
+/* evidence chart gallery */
+.chart-section{max-width:1180px;margin:0 auto}
+.chart-grid{display:grid;grid-template-columns:repeat(3,1fr);gap:16px}
+.chart-card{
+  display:flex;flex-direction:column;gap:14px;text-decoration:none;color:inherit;
+  padding:16px;transition:all .25s;
+}
+.chart-card:hover{border-color:rgba(56,189,248,.24);transform:translateY(-2px)}
+.chart-meta{display:flex;gap:10px;align-items:flex-start;min-height:64px}
+.chart-meta svg{color:var(--accent3);flex-shrink:0;margin-top:2px}
+.chart-meta h4{font-size:14px;font-family:var(--display);margin-bottom:5px}
+.chart-meta p{font-size:12px;line-height:1.45;color:var(--ink2)}
+.chart-card img{
+  display:block;width:100%;aspect-ratio:5/3;object-fit:cover;
+  border-radius:10px;border:1px solid rgba(255,255,255,.06);
+  background:#f8fafc;
+}
+
 /* ══════════════════════════════════════════════════════
    MISSION CONTROL
    ══════════════════════════════════════════════════════ */
@@ -365,13 +383,14 @@ button,select,input{font:inherit;color:inherit}
 @media(max-width:1024px){
   .mc,.ba-grid,.jw-inline{grid-template-columns:1fr}
   .arch-grid{grid-template-columns:repeat(2,1fr)}
+  .chart-grid{grid-template-columns:repeat(2,1fr)}
   .score-strip{grid-template-columns:repeat(2,1fr)}
 }
 @media(max-width:640px){
   .hdr{padding:0 14px;gap:10px;height:52px}
   .hdr-nav{display:none}
   .main{padding:14px 14px 40px}
-  .arch-grid,.ac-grid,.ac-auto,.brief-grid,.jw-results,.jw-compare{grid-template-columns:1fr}
+  .arch-grid,.chart-grid,.ac-grid,.ac-auto,.brief-grid,.jw-results,.jw-compare{grid-template-columns:1fr}
   .score-strip{grid-template-columns:1fr}
   .land-hero{padding:48px 12px 32px}
   .land-hero h1{font-size:36px}