Align leaderboard with paper core protocol

README.md

@@ -22,4 +22,4 @@ The Space reads leaderboard tables from the dataset repository:
 
 The dataset repository remains the source of record. This Space is only the web presentation layer.
 
-The default view is paper-first: it anchors the web leaderboard to the camera-ready Table 2 / Figure 3 six-family summary. The 22-row neural-style board is retained as a post-rebuttal expansion view rather than as the paper's primary leaderboard.
+The default view is paper-first: it anchors the web leaderboard to the camera-ready Table 2 / Figure 3 six-family summary from the 6-scenario, 50-draw core protocol. Table 2 uses tier-balanced aggregation over valid metric values: stationary columns average Tier 1 and Tier 2 means with equal weight, and non-stationary columns use Tier 3 means. The 22-method synthetic and semi-synthetic views are benchmark-only living extensions that use the released five-metric ranking files from the dataset repository; they are not replacements for the primary Table 2 view and are not camera-ready paper claims.

app.py

@@ -18,8 +18,10 @@ SITE_FILES = {
     "paper_global": "data/paper_tables/global_performance_summary.csv",
     "paper_radar": "data/paper_figures/selected_radar_charts.png",
     "synthetic": f"site_data/{DATA_VERSION}/leaderboard_synthetic_full22_overall.csv",
+    "synthetic_rank": "data/synthetic_full22_extension/results/summary/ranking_paper_5metric_overall.csv",
     "real_proxy": f"site_data/{DATA_VERSION}/leaderboard_real_proxy22_overall.csv",
     "semisynth": f"site_data/{DATA_VERSION}/leaderboard_semisynth_transfer_by_method.csv",
+    "semisynth_rank": "data/semisynth_transfer/results/summary/ranking_paper_5metric_overall.csv",
     "track_b": f"site_data/{DATA_VERSION}/leaderboard_post_rebuttal_real_physics_track_b.csv",
     "track_c": f"site_data/{DATA_VERSION}/leaderboard_post_rebuttal_real_proxy_track_c.csv",
     "methods": f"site_data/{DATA_VERSION}/methods.json",
@@ -29,9 +31,9 @@ SITE_FILES = {
 
 VIEW_LABELS = {
     "paper_global": "Paper Table 2 / Figure 3 Anchors",
-    "synthetic": "Post-Rebuttal Synthetic Full22 Expansion",
-    "real_proxy": "Paper-Aligned Real Proxy 22",
-    "semisynth": "Semi-Synthetic Transfer",
+    "synthetic": "Benchmark 22-Method Synthetic Extension",
+    "real_proxy": "Benchmark 22-Method Real Proxy",
+    "semisynth": "Benchmark 22-Method Semi-Synthetic Transfer",
     "track_b": "Post-Rebuttal Track B: CO2 / Tides",
     "track_c": "Post-Rebuttal Track C: Real Proxy Panel",
 }
@@ -39,25 +41,30 @@ VIEW_LABELS = {
 VIEW_NOTES = {
     "paper_global": (
         "Camera-ready paper-aligned view. This reproduces Table 2's six-family "
-        "global performance summary. Table 2 reports two of the five paper-core "
-        "component-recovery metrics, split by stationary and non-stationary regimes; "
-        "Figure 3 shows the five-axis capability profile. This is not a single "
-        "universal ranking."
+        "global performance summary from the 6-scenario, 50-draw core protocol. "
+        "Table 2 reports two of the five paper-core component-recovery metrics, "
+        "using tier-balanced stationary aggregation over Tiers 1-2 and Tier 3 for "
+        "the non-stationary columns; Figure 3 shows the "
+        "five-axis capability profile. This is not a single universal ranking."
     ),
     "synthetic": (
-        "Post-rebuttal expanded synthetic board: 6 scenarios x 50 draws = 300 generated "
+        "Benchmark-only expanded synthetic board: 6 scenarios x 50 draws = 300 generated "
         "time-series instances. The expanded roster gives 22 method rows, "
-        "or 6,600 method-instance evaluations. This is an extension view, not the "
+        "or 6,600 method-instance evaluations. This is a living extension view, not the "
         "paper's primary Table 2/Figure 3 view. It reuses the five component-recovery "
-        "metrics and adds a mean-rank convenience summary."
+        "metrics and uses the released five-metric mean within-setting ranking."
     ),
     "real_proxy": (
-        "Paper-aligned real-data companion track. These are proxy diagnostics, "
-        "not exact component-recovery labels. Lower overall mean rank is better."
+        "Benchmark-only real-data companion track. These are proxy diagnostics, "
+        "not exact component-recovery labels and not camera-ready paper rankings. "
+        "Lower overall mean rank is better."
     ),
     "semisynth": (
-        "Known synthetic mechanisms injected into real monthly backgrounds. "
-        "The displayed rank is recomputed from trend and seasonal recovery metrics."
+        "Benchmark-only 22-method semi-synthetic transfer track: known synthetic "
+        "mechanisms injected into six real monthly backgrounds with three mechanisms, "
+        "two background scales, and eight windows per setting. This is a living "
+        "benchmark extension, not a camera-ready paper table; the displayed rank uses "
+        "the released five-metric mean within-setting ranking."
     ),
     "track_b": (
         "Post-rebuttal mechanism-aware checks on CO2 and tides. These use "
@@ -112,9 +119,9 @@ HIGHER_IS_BETTER = {
 
 DEFAULT_METRIC = {
     "paper_global": "stationary_trend_r2",
-    "synthetic": "mean_metric_rank",
+    "synthetic": "overall_mean_rank",
     "real_proxy": "overall_mean_rank",
-    "semisynth": "mean_metric_rank",
+    "semisynth": "overall_mean_rank",
     "track_b": "seasonal_r2",
     "track_c": "band_plausibility",
 }
@@ -136,6 +143,7 @@ COLUMN_LABELS = {
     "metric_S_maxlag_corr_mean": "Seasonal max-lag corr mean",
     "mean_metric_rank": "Mean metric rank",
     "overall_mean_rank": "Overall mean rank",
+    "overall_std_rank": "Overall rank SD",
     "coverage_success": "Coverage",
     "row_count": "Rows",
     "setting_count": "Settings",
@@ -154,6 +162,8 @@ TABLE_LABELS = {
     "metric_S_r2_mean": "S R2",
     "metric_S_spectral_corr_mean": "S spec.",
     "metric_S_maxlag_corr_mean": "S max-lag",
+    "overall_mean_rank": "Mean rank",
+    "overall_std_rank": "Rank SD",
 }
 
 PAPER_GLOBAL_METRICS = [
@@ -231,7 +241,8 @@ def metrics_html() -> str:
       <p>
         <strong>Table 2</strong> displays only <strong>{int(meta.get("paper_table2_display_metric_count", 2))}</strong>
         of those metrics: Trend R2 and Seasonal spectral correlation, split over stationary regimes (Tiers 1-2)
-        and the non-stationary regime (Tier 3). <strong>Figure 3</strong> is the five-metric profile view.
+        and the non-stationary regime (Tier 3). Its stationary columns use a tier-balanced average of Tier 1
+        and Tier 2 means over valid metric values; non-stationary columns use Tier 3 means. <strong>Figure 3</strong> is the five-metric profile view.
       </p>
       <div class="metric-grid">{''.join(cards)}</div>
     </section>
@@ -290,6 +301,24 @@ def _add_mean_rank(df: pd.DataFrame, metrics: list[str]) -> pd.DataFrame:
     return out
 
 
+def _attach_released_rank(df: pd.DataFrame, rank_file: str) -> pd.DataFrame:
+    rank = pd.read_csv(_download(rank_file))
+    rank = _numeric(rank)
+    keep = [
+        col
+        for col in ["method", "overall_mean_rank", "overall_std_rank", "setting_count", "rank_order"]
+        if col in rank.columns
+    ]
+    out = df.drop(columns=[col for col in keep if col != "method" and col in df.columns])
+    out = out.merge(rank[keep], on="method", how="left")
+    if "rank_order" in out.columns:
+        out["rank"] = pd.to_numeric(out["rank_order"], errors="coerce")
+    elif "overall_mean_rank" in out.columns:
+        values = pd.to_numeric(out["overall_mean_rank"], errors="coerce")
+        out["rank"] = values.rank(ascending=True, method="min")
+    return out
+
+
 @lru_cache(maxsize=1)
 def load_data() -> dict[str, pd.DataFrame]:
     methods = _method_table()
@@ -297,16 +326,7 @@ def load_data() -> dict[str, pd.DataFrame]:
 
     synthetic = pd.read_csv(_download(SITE_FILES["synthetic"]))
     synthetic = _attach_methods(_numeric(synthetic), methods)
-    synthetic = _add_mean_rank(
-        synthetic,
-        [
-            "metric_T_r2_mean",
-            "metric_T_dtw_mean",
-            "metric_S_r2_mean",
-            "metric_S_spectral_corr_mean",
-            "metric_S_maxlag_corr_mean",
-        ],
-    )
+    synthetic = _attach_released_rank(synthetic, SITE_FILES["synthetic_rank"])
 
     real_proxy = pd.read_csv(_download(SITE_FILES["real_proxy"]))
     real_proxy = _attach_methods(_numeric(real_proxy), methods)
@@ -314,16 +334,7 @@ def load_data() -> dict[str, pd.DataFrame]:
 
     semisynth = pd.read_csv(_download(SITE_FILES["semisynth"]))
     semisynth = _attach_methods(_numeric(semisynth), methods)
-    semisynth = _add_mean_rank(
-        semisynth,
-        [
-            "metric_T_r2_mean",
-            "metric_T_dtw_mean",
-            "metric_S_r2_mean",
-            "metric_S_spectral_corr_mean",
-            "metric_S_maxlag_corr_mean",
-        ],
-    )
+    semisynth = _attach_released_rank(semisynth, SITE_FILES["semisynth_rank"])
 
     track_b = pd.read_csv(_download(SITE_FILES["track_b"]))
     track_b = _attach_methods(_numeric(track_b), methods)
@@ -373,7 +384,10 @@ def filter_table(view: str, metric: str, family: str, query: str, top_n: int):
         df = df[mask]
 
     ascending = not HIGHER_IS_BETTER.get(metric, True)
-    ranked = df.sort_values(metric, ascending=ascending, na_position="last").head(int(top_n))
+    if view == "paper_global":
+        ranked = df.head(int(top_n))
+    else:
+        ranked = df.sort_values(metric, ascending=ascending, na_position="last").head(int(top_n))
 
     if view == "paper_global":
         show_cols = ["display_name", metric]