update

src/similarity_model/hybrid_ranker.py

@@ -134,20 +134,24 @@ def compute_originality(
 ) -> float:
     """
     Originality Score (0-100)
-
-    - Driven strictly by inverse of hybrid similarity (non-linear).
+    Calibrated piecewise curve to match human perception of similarity.
     """
-
     hybrid_score = clamp(hybrid_score)
 
-    # Non-linear penalty: 50% similarity drops originality slightly (to 75%), 
-    # but 90% similarity drops it massively (to 19%).
-    originality = 1.0 - (hybrid_score ** 2)
-
-    return round(
-        max(0.0, min(100.0, originality * 100)),
-        2
-    )
+    if hybrid_score >= 0.39:
+        # > 0.39 is a strong semantic match -> Originality 0-25%
+        originality = 25.0 - ((hybrid_score - 0.39) / 0.11) * 25.0
+    elif hybrid_score >= 0.35:
+        # 0.35 to 0.39 is a moderate match -> Originality 25-75%
+        originality = 75.0 - ((hybrid_score - 0.35) / 0.04) * 50.0
+    elif hybrid_score >= 0.25:
+        # 0.25 to 0.35 is a weak match -> Originality 75-95%
+        originality = 95.0 - ((hybrid_score - 0.25) / 0.10) * 20.0
+    else:
+        # < 0.25 is effectively unrelated -> Originality 95-100%
+        originality = 100.0 - (hybrid_score / 0.25) * 5.0
+
+    return round(clamp(originality), 2)
 
 
 def compute_confidence(