Align hard split IDs with published ranking

metadata/splits/hard.json

@@ -13,8 +13,8 @@
     "armbruster2012",
     "watermeyer2020",
     "zohali2022",
-    "wei2020",
     "legrain2020",
+    "wei2020",
     "roshanaei2017",
     "adulyasak2015",
     "desaulniers2010",
@@ -47,8 +47,8 @@
     "roberti2018",
     "kobeaga2024",
     "bierwirth2017",
-    "bard2002",
     "fragkos2021",
+    "bard2002",
     "delage2022",
     "wu2018",
     "spliet2015",

metadata/splits/hard_methodology.md

@@ -6,8 +6,7 @@ reference performance.
 
 ## Published evidence
 
-- `hard.json`: authoritative selected task IDs. Its order is for presentation;
-  membership, not order, is the selection invariant.
+- `hard.json`: authoritative selected task IDs in descending hard-rank order.
 - `hard_selection_inputs_180.csv`: frozen, one-row-per-task input snapshot used
   for this selection. It includes the relevant paper metadata, mean instance
   size fields, and historical five-large-instance Gurobi summary statistics.
@@ -85,9 +84,5 @@ To materialize the computed ranking for comparison:
 python metadata/splits/reproduce_hard.py --ranking-out /tmp/hard_ranking_180.csv
 ```
 
-The script exits nonzero if the computed top-50 membership differs from
-`hard.json`. The published input snapshot has SHA-256
-`90902079c17a202508255df756f43c9f0bee29b49076141a0ceb83c8161fab50`.
-The recomputed 180-task ordering and scores match the original benchmark
-ranking exactly. `hard.json` has the same 50 members but swaps two pairs of
-adjacent IDs relative to ranking order; its ordering is not a score rank.
+The script verifies that the computed top 50 IDs and their ranking order
+match `hard.json`.

metadata/splits/reproduce_hard.py

@@ -113,13 +113,13 @@ def main() -> int:
         ]
         ranking["selected_hard"] = ranking["paper_id"].isin(selected)
         ranking[columns].to_csv(args.ranking_out, index=False)
-    computed = set(ranking.head(hard["task_count"])["paper_id"])
-    if computed != selected:
-        print(f"Selection differs: computed-only={sorted(computed - selected)}, "
-              f"published-only={sorted(selected - computed)}")
+    computed = ranking.head(hard["task_count"])["paper_id"].tolist()
+    if computed != hard["task_ids"]:
+        print(f"Selection or order differs: computed-only={sorted(set(computed) - selected)}, "
+              f"published-only={sorted(selected - set(computed))}")
         return 1
     print(f"Reproduced {len(selected)} hard tasks from {len(ranking)} frozen input rows.")
-    print("Membership matches hard.json; its task_ids order is authoritative for display.")
+    print("Task IDs and ranking order match hard.json.")
     return 0