sync(report): brax_hv_bestreg72_refm5000m7000_ep16_20260714

reports/v0.52.3.0/brax_hv_bestreg72_refm5000m7000_ep16_20260713/README.md

@@ -64,7 +64,12 @@ B50 configuration has the tightest three-seed spread among the leaders.
 - `analysis/overview_metrics.json`: machine-readable headline metrics.
 - `analysis/plots/`: compact overview, top-10 chart, and 12 detailed plots.
 - `calibration/`: common-reference calibration and 216-row mesh check.
-- `make_overview.py`: reproducible generator for the two compact overview plots.
+- `make_overview.py`: reproducible generator for the compact and overlay plots.
+
+In the nine detailed `(H, flip)` budget plots, A0 and A1 share the same axes.
+Solid circles denote A0, dashed squares denote A1, faint points show the three
+individual seeds, and each translucent band is **mean +/- population SD**. The
+band is descriptive uncertainty for three seeds, not a 95% confidence interval.
 
 The full policy-bank payload is published separately in the configured model
 repository as a compressed archive with a SHA-256 sidecar and an explicit

reports/v0.52.3.0/brax_hv_bestreg72_refm5000m7000_ep16_20260713/make_overview.py

@@ -47,6 +47,98 @@ def variant_effect(
     }
 
 
+def plot_detailed_overlays(
+    runs: list[dict[str, str]],
+    variants: list[str],
+    colors: dict[str, str],
+    pretty: dict[str, str],
+) -> list[Path]:
+    groups: dict[
+        tuple[str, str], dict[str, dict[str, list[float]]]
+    ] = defaultdict(lambda: defaultdict(lambda: defaultdict(list)))
+    for row in runs:
+        key = (row["H"], row["flip"])
+        groups[key][row["variant"]][row["B"]].append(
+            float(row["hypervolume"]) / 1e6
+        )
+
+    markers = {"a0_resmlp": "o", "a1_resmlp": "s"}
+    linestyles = {"a0_resmlp": "-", "a1_resmlp": "--"}
+    offsets = {"a0_resmlp": -0.035, "a1_resmlp": 0.035}
+    seed_jitter = [-0.012, 0.0, 0.012]
+    output_paths: list[Path] = []
+
+    for (horizon, flip), by_variant in sorted(
+        groups.items(), key=lambda item: (int(item[0][0]), item[0][1])
+    ):
+        budgets = sorted(
+            {budget for values in by_variant.values() for budget in values},
+            key=int,
+        )
+        x_base = list(range(len(budgets)))
+        fig, ax = plt.subplots(figsize=(10, 6), constrained_layout=True)
+
+        for variant in variants:
+            x = [value + offsets[variant] for value in x_base]
+            values = [by_variant[variant][budget] for budget in budgets]
+            means = [mean(item) for item in values]
+            stds = [pstdev(item) for item in values]
+            lower = [center - spread for center, spread in zip(means, stds)]
+            upper = [center + spread for center, spread in zip(means, stds)]
+
+            ax.fill_between(
+                x,
+                lower,
+                upper,
+                color=colors[variant],
+                alpha=0.14,
+                linewidth=0,
+            )
+            ax.plot(
+                x,
+                means,
+                color=colors[variant],
+                marker=markers[variant],
+                linestyle=linestyles[variant],
+                linewidth=2.4,
+                markersize=7,
+                label=f"{pretty[variant]} mean +/- SD",
+            )
+            for budget_index, seed_values in enumerate(values):
+                for seed_index, value in enumerate(seed_values):
+                    jitter = seed_jitter[seed_index] if seed_index < 3 else 0.0
+                    ax.scatter(
+                        budget_index + offsets[variant] + jitter,
+                        value,
+                        color=colors[variant],
+                        alpha=0.42,
+                        s=20,
+                        edgecolors="none",
+                        zorder=3,
+                    )
+
+        flip_label = flip.replace("p", ".")
+        ax.set_title(
+            f"HalfCheetah HV | H={horizon} | flip={flip_label} | "
+            "coverage=1.3 | ref=(-5000, -7000)"
+        )
+        ax.set_xlabel("Budget B")
+        ax.set_ylabel("Hypervolume (million)")
+        ax.set_xticks(x_base, budgets)
+        ax.grid(alpha=0.25)
+        ax.legend(frameon=False, title="3 seeds per point")
+
+        output_path = PLOTS / (
+            "hv_by_budget_mo-halfcheetah-v5_braxvmap_"
+            f"H{horizon}_flip{flip}_refm5000_m7000.png"
+        )
+        fig.savefig(output_path, dpi=180)
+        plt.close(fig)
+        output_paths.append(output_path)
+
+    return output_paths
+
+
 def main() -> None:
     PLOTS.mkdir(parents=True, exist_ok=True)
     runs = read_tsv(ANALYSIS / "parsed_hv_runs.tsv")
@@ -56,6 +148,8 @@ def main() -> None:
     colors = {"a0_resmlp": "#2878B5", "a1_resmlp": "#E07B39"}
     pretty = {"a0_resmlp": "A0", "a1_resmlp": "A1"}
 
+    detailed_paths = plot_detailed_overlays(runs, variants, colors, pretty)
+
     by_variant: dict[str, list[float]] = defaultdict(list)
     for row in runs:
         by_variant[row["variant"]].append(float(row["hypervolume"]))
@@ -194,6 +288,7 @@ def main() -> None:
     )
     print(f"overview_plot={overview_path}")
     print(f"top10_plot={PLOTS / 'top10_hv_configs.png'}")
+    print(f"detailed_overlay_plots={len(detailed_paths)}")
     print(f"metrics={ANALYSIS / 'overview_metrics.json'}")