first commit

apps/inspector/test_semantic_real.py

@@ -0,0 +1,47 @@
+"""Opt-in operational fixtures using staged real Google weights, CPU only."""
+import io
+import math
+import os
+import unittest
+import wave
+from engine import get_model,embedding_features,inspect_content,cosine
+from semantic import validate_vectors,DIMENSIONS
+
+@unittest.skipUnless(os.environ.get('SAFETYNET_TEST_NATIVE_SEMANTIC')=='1','Opt in to verified native CPU fixtures')
+class NativeCPUFixtures(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):cls.model=get_model()
+    @classmethod
+    def tearDownClass(cls):cls.model.close()
+    def test_real_paraphrase_blocks_and_unrelated_text_allows(self):
+        reference=embedding_features(b'Our confidential launch plan schedules the new product announcement for next October.')
+        protected=b'The secret product rollout will be announced in October next year.'
+        public=b'The public cafeteria menu includes tomato soup and bread.'
+        related=embedding_features(protected);unrelated=embedding_features(public)
+        self.assertGreater(cosine(reference['vector'],related['vector']),.8)
+        self.assertLess(cosine(reference['vector'],unrelated['vector']),.8)
+        policy={'dlp':{'maxBytes':1048576,'rules':[],'semantic':{**self.model.info(),'enabled':True,'categories':[{'id':'native-reference','prototype':reference['vector'],'prototypes':reference['vectors'],'threshold':.8}]}}}
+        self.assertEqual(inspect_content(protected,policy)['ruleId'],'native-reference')
+        self.assertEqual(inspect_content(public,policy)['decision'],'allow')
+    def test_native_token_windows_cover_long_document_tail(self):
+        text='Public synthetic office lunch menu and team meeting schedule. '*600+'Final unique coverage marker.'
+        result=embedding_features(text.encode())
+        self.assertGreater(result['coverage']['textChunks'],1)
+        self.assertTrue(result['coverage']['complete']);validate_vectors(result['vectors'])
+    def test_native_image_encoder_returns_real_normalized_vector(self):
+        from PIL import Image,ImageDraw
+        image=Image.new('RGB',(192,192),'white');draw=ImageDraw.Draw(image)
+        draw.rectangle((20,20,90,170),fill='blue');draw.ellipse((110,50,175,115),fill='red')
+        output=io.BytesIO();image.save(output,format='PNG')
+        vectors=self.model.image_vectors(output.getvalue())
+        self.assertEqual(len(vectors),1);validate_vectors(vectors)
+    def test_native_audio_encoder_returns_real_normalized_vector(self):
+        import numpy as np
+        samples=(np.sin(np.arange(16000)*2*math.pi*440/16000)*12000).astype('<i2')
+        output=io.BytesIO()
+        with wave.open(output,'wb') as stream:
+            stream.setnchannels(1);stream.setsampwidth(2);stream.setframerate(16000);stream.writeframes(samples.tobytes())
+        vectors=self.model.audio_vectors(output.getvalue())
+        self.assertEqual(len(vectors),1);validate_vectors(vectors)
+
+if __name__=='__main__':unittest.main()