feat: updated the Croissant metadata

croissant.json

@@ -684,8 +684,42 @@
             "description": "English Wikipedia Snapshot (January 1, 2023)"
         }
     ],
-    "prov:wasGeneratedBy": {
-        "@type": "prov:Activity",
-        "description": "The dataset was generated through a multi-stage pipeline: (1) Breadth-First Search (BFS) Wikipedia category traversal, (2) Named Entity Recognition (NER) for factual density filtering, (3) Lexical and syntactic anonymization of pronouns and entity names, (4) Generational Proximity Filtering for distractor matching, and (5) a human-in-the-loop audit for bijective mapping and factual accuracy."
-    }
+    "prov:wasGeneratedBy": [
+        {
+            "@type": "prov:Activity",
+            "name": "Entity Discovery and Stratification",
+            "description": "Automated Breadth-First Search (BFS) traversal of English Wikipedia category graphs to identify candidate entities. Entities were ranked and stratified into Upper and Lower tiers using a multimodal popularity score (Sp).",
+            "wasAssociatedWith": {
+                "@type": "prov:SoftwareAgent",
+                "name": "ShadowBench Wiki-Crawler"
+            }
+        },
+        {
+            "@type": "prov:Activity",
+            "name": "Factual Mining and Hardening (v1-v3)",
+            "description": "Extraction of factual traits from the January 1, 2023 Wikipedia snapshot. Applied spaCy-based NER filtering for attribute density. Implemented an iterative hardening pipeline including lexical anonymization, pronoun neutralization, and the application of a 25-year Generational Proximity Filter (GPF).",
+            "wasAssociatedWith": {
+                "@type": "prov:SoftwareAgent",
+                "name": "spaCy NER and Anonymization Pipeline"
+            }
+        },
+        {
+            "@type": "prov:Activity",
+            "name": "Adversarial MCQ Synthesis",
+            "description": "Combinatorial permutation of traits (Trait A to Trait B) with gender-homogeneous hard-negative distractor matching to prevent non-semantic shortcut learning.",
+            "wasAssociatedWith": {
+                "@type": "prov:SoftwareAgent",
+                "name": "ShadowBench MCQ Generator"
+            }
+        },
+        {
+            "@type": "prov:Activity",
+            "name": "Human-in-the-loop Audit",
+            "description": "Manual quality review by the research team to ensure bijective mapping (uniqueness) of shadow descriptions and factual accuracy across all 7,000+ QA pairs.",
+            "wasAssociatedWith": {
+                "@type": "prov:Person",
+                "name": "ShadowBench Research Team"
+            }
+        }
+    ]
 }