Clarify overall tables and tracked rank history

app.py

@@ -313,7 +313,7 @@ with demo:
                 elem_classes="markdown-text",
             )
             gr.Markdown(
-                "**Cumulative live evaluation.** MSE, RMSE, MAPE, CRPS, and Stability are lower-is-better; RTG is higher-is-better; more negative Improvement indicates a stronger decreasing-error trend.",
+                "**Live cumulative results (release-history aggregate).** Uses every real release collected up to now. MSE, RMSE, MAPE, CRPS, and Stability are lower-is-better; RTG is higher-is-better; more negative Improvement means errors decrease more consistently over releases.",
                 elem_classes="markdown-text",
             )
             live_overall_df = _leaderboard_dataframe(
@@ -321,7 +321,7 @@ with demo:
                 label="Up-to-now overall results",
             )
             gr.Markdown(
-                "**Average Rank, Win Rate, and Elo**",
+                "**Live relative ranks (same release history).** Average Rank, Win Rate, and Elo are computed from model-vs-model comparisons on shared live releases, so they summarize relative standing rather than absolute error values.",
                 elem_classes="markdown-text",
             )
             live_rank_df = gr.HTML(
@@ -341,20 +341,24 @@ with demo:
                 label="Domain distribution",
             )
             gr.Markdown(
-                "Geometric mean of **absolute** metrics across all datasets. MSE/CRPS are lower-is-better; RankScore is higher-is-better.",
+                "**Latest snapshot aggregate (not release-history).** Geometric mean of absolute metric values across the current dataset snapshot. MSE/CRPS are lower-is-better; RankScore is higher-is-better. This table is comparable to a static leaderboard snapshot, while the live tables above use temporal release histories.",
                 elem_classes="markdown-text",
             )
             overall_values_df = _leaderboard_dataframe(
                 value=INITIAL_OUTPUTS[2],
                 label="Metric values and RankScore",
             )
+            gr.Markdown(
+                "**Latest snapshot ranks.** Per-metric ranks derived from the snapshot values above; lower rank is better.",
+                elem_classes="markdown-text",
+            )
             overall_ranks_df = gr.HTML(
                 value=INITIAL_OUTPUTS[3],
                 elem_classes="rank-html-output",
             )
 
             gr.Markdown(
-                "**Baseline daily rank history** · average rank across all datasets per day (lower is better)",
+                "**Tracked model daily rank history** · the Model column names the model tracked on each date; rank values are its average per-dataset ranks for that day (lower is better).",
                 elem_classes="markdown-text",
             )
             baseline_history_df = gr.HTML(
@@ -362,7 +366,7 @@ with demo:
                 elem_classes="rank-html-output",
             )
             gr.Markdown(
-                "**Baseline weekly rank history** · average of daily ranks per ISO week (lower is better)",
+                "**Tracked model weekly rank history** · grouped by ISO week and Model; values are weekly means of the daily tracked-model ranks (lower is better).",
                 elem_classes="markdown-text",
             )
             baseline_weekly_df = gr.HTML(
@@ -370,7 +374,7 @@ with demo:
                 elem_classes="rank-html-output",
             )
             gr.Markdown(
-                "**Baseline rank trend** · MSE / CRPS / MASE / MAE rank over time",
+                "**Tracked model rank trend** · MSE / CRPS / MASE / MAE rank over time; see the table Model column for which model each point tracks.",
                 elem_classes="markdown-text",
             )
             rank_trend_plot = gr.Plot(