Upload report_flags.py

report_flags.py

@@ -156,7 +156,7 @@ def read_score(text):
     """(percent, kind, tool) read from the report's first pages, e.g. (23.0, 'AI', 'Turnitin').
     kind is 'AI' or 'Similarity'; percent is None if no score could be found."""
     low = (text or "").lower()
-    tool = next((name for key, name in _TOOLS if key in low), "Your report")
+    tool = next((name for key, name in _TOOLS if key in low), "Report")
     best = None                                        # (priority, position, pct, kind)
     for m in _PCT.finditer(low):
         pct = float(m.group(1))
@@ -185,6 +185,9 @@ def read_score(text):
     return best[2], best[3], tool
 
 
+KEEP_OVERLAP = 0.5      # wording overlap at/above which a flagged sentence is assumed to stay flagged
+
+
 def estimate_after(score, pairs):
     """Estimated report score after rewriting. pairs = [(original flagged sentence, final sentence)].
     The report's score comes from its flagged text; each flagged sentence keeps the share of its wording that is
@@ -196,7 +199,10 @@ def estimate_after(score, pairs):
     for o, n in pairs:
         w = max(1, len(o.split()))
         tot += w
-        kept += w * (1.0 if n.strip() == o.strip() else similarity_to_source(o, n) / 100.0)
+        # detectors also flag lightly reworded (AI-paraphrased) text: a sentence only starts to count as cleared
+        # once more than half of its wording has changed (overlap below 50%), reaching zero at 0% overlap
+        ov = 1.0 if n.strip() == o.strip() else similarity_to_source(o, n) / 100.0
+        kept += w * min(1.0, ov / KEEP_OVERLAP)
     return round(score * kept / tot, 1) if tot else score