Stronger AI detector: paraphrase-robust third model in the ensemble

fakespot-ai/roberta-base-ai-text-detection-v1 averaged with the current coai + desklib score. Measured
(devtools/compare_detectors.py): AI samples 18/18 flagged, human abstracts from 2008 0/21, paragraphs our own
humanizer had brought to 6%: 6/16 flagged (was 1/16). RADAR was tested and rejected (separation 0.72 on academic
text). DETECTOR_MODEL3="" turns it off; DETECTOR_WEIGHT3 sets its share (0.5).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

detector.py

@@ -54,6 +54,42 @@ WEIGHT2 = float(os.getenv("DETECTOR_WEIGHT2", "0.7"))
 COMBINE = os.getenv("DETECTOR_COMBINE", "mean").strip().lower()
 _tok2, _model2, _m2_failed = None, None, False
 
+# Third detector: robust to paraphrasing (what Turnitin's "AI-paraphrased" check is about). Measured with
+# devtools/compare_detectors.py on 18 AI-written academic samples, 21 human abstracts from 2008 and 16 paragraphs our
+# own humanizer had brought to 6%: coai + desklib flagged 18/18 AI, 0/21 human, 1/16 humanized; this RoBERTa
+# (fakespot-ai) flagged 18/18, 2/21, 8/16. Averaged with the first two: 18/18 AI, 0/21 human, 6/16 humanized, so
+# the score is much harder to game without flagging human writing.
+#   DETECTOR_MODEL3   default fakespot-ai/roberta-base-ai-text-detection-v1 ("" = off)
+#   DETECTOR_WEIGHT3  default 0.5   share of the final score
+MODEL3_ID = os.getenv("DETECTOR_MODEL3", "fakespot-ai/roberta-base-ai-text-detection-v1").strip()
+WEIGHT3 = float(os.getenv("DETECTOR_WEIGHT3", "0.5"))
+_tok3, _model3, _m3_failed = None, None, False
+_cpu_m3 = None
+
+
+def load_model3():
+    global _tok3, _model3, _m3_failed
+    if not MODEL3_ID or _m3_failed or _model3 is not None:
+        return _tok3, _model3
+    try:
+        import torch
+        from transformers import AutoModelForSequenceClassification, AutoTokenizer
+        _tok3 = AutoTokenizer.from_pretrained(MODEL3_ID, token=HF_TOKEN)
+        _model3 = AutoModelForSequenceClassification.from_pretrained(MODEL3_ID, token=HF_TOKEN).eval()
+        if ZERO or torch.cuda.is_available():
+            _model3.to("cuda")
+        print(f"[detector] third model loaded: {MODEL3_ID}", flush=True)
+    except Exception as err:
+        _m3_failed = True
+        print(f"[detector] third model not available ({err}); first two only", flush=True)
+    return _tok3, _model3
+
+
+def _with3(p12, p3):
+    if p3 is None:
+        return p12
+    return [(1 - WEIGHT3) * a + WEIGHT3 * b for a, b in zip(p12, p3)]
+
 
 def _desklib_class():
     import torch
@@ -139,6 +175,7 @@ def load_model():
 if os.getenv("SKIP_MODEL_LOAD") != "1":
     load_model()
     load_model2()
+    load_model3()
     DX.start_lo_init()            # fetch newest LibreOffice in the background
 
 def _score_with(tok, model, texts):
@@ -162,7 +199,9 @@ def _score_gpu(texts):
     p1 = _score_with(tok, model, texts)
     tok2, model2 = load_model2()
     p2 = _score_with2(tok2, model2, texts) if model2 is not None else None
-    return _combine(p1, p2)
+    tok3, model3 = load_model3()
+    p3 = _score_with(tok3, model3, texts) if model3 is not None else None   # label 1 = "ai"
+    return _with3(_combine(p1, p2), p3)
 
 
 _cpu_det, _cpu_det_lock = None, __import__("threading").Lock()
@@ -170,7 +209,7 @@ LAST_ENGINE = "GPU"
 
 
 def _score_cpu(texts):
-    global _cpu_det
+    global _cpu_det, _cpu_m3
     import torch
     from transformers import AutoModelForSequenceClassification
     with _cpu_det_lock:
@@ -185,7 +224,17 @@ def _score_cpu(texts):
             if next(model2.parameters()).device.type != "cpu":
                 model2 = _cpu_model2()
             p2 = _score_with2(tok2, model2, texts)
-        return _combine(p1, p2)
+        p3 = None
+        tok3, model3 = load_model3()
+        if model3 is not None:
+            if next(model3.parameters()).device.type != "cpu":
+                with _cpu_det_lock:
+                    if _cpu_m3 is None:
+                        from transformers import AutoModelForSequenceClassification as _A
+                        _cpu_m3 = _A.from_pretrained(MODEL3_ID, token=HF_TOKEN).eval()
+                model3 = _cpu_m3
+            p3 = _score_with(tok3, model3, texts)
+        return _with3(_combine(p1, p2), p3)
 
 
 _cpu_m2 = None

devtools/compare_detectors.py

@@ -60,6 +60,8 @@ def main():
         scorers["current (live: coai + desklib)"] = D.score_texts
     if "radar" in want:
         scorers["RADAR (paraphrase-robust)"] = hf_scorer("TrustSafeAI/RADAR-Vicuna-7B", ai_index=0)
+    if "superannotate" in want:
+        scorers["SuperAnnotate"] = hf_scorer("SuperAnnotate/ai-detector")
     if "fakespot" in want:
         scorers["fakespot roberta"] = hf_scorer("fakespot-ai/roberta-base-ai-text-detection-v1")
     if "desklib" in want: