Add a predicate that finds endpoints with logically-inconsistent characteristics, and exclude such endpoints from both positive and negative examples extracted for the codex prompt.

java/ql/experimental/adaptivethreatmodeling/lib/experimental/adaptivethreatmodeling/EndpointCharacteristics.qll

@@ -16,6 +16,51 @@ private import experimental.adaptivethreatmodeling.RequestForgeryATM
 private import semmle.code.java.security.ExternalAPIs as ExternalAPIs
 private import semmle.code.java.Expr as Expr
 
+/**
+ * Holds if the given endpoint has a self-contradictory combination of characteristics. Detects errors in our endpoint
+ * characteristics. Lists the problematic characterisitics and their implications for all such endpoints, together with
+ * an error message indicating why this combination is problematic.
+ *
+ * Copied from javascript/ql/experimental/adaptivethreatmodeling/test/endpoint_large_scale/ContradictoryEndpointCharacteristics.ql
+ */
+query predicate erroneousEndpoints(
+  DataFlow::Node endpoint, EndpointCharacteristic characteristic, EndpointType endpointClass,
+  float confidence, string errorMessage
+) {
+  // An endpoint's characteristics should not include positive indicators with medium/high confidence for more than one
+  // class.
+  exists(EndpointCharacteristic characteristic2, EndpointType endpointClass2, float confidence2 |
+    endpointClass.getEncoding() != endpointClass2.getEncoding() and
+    characteristic.appliesToEndpoint(endpoint) and
+    characteristic2.appliesToEndpoint(endpoint) and
+    characteristic.hasImplications(endpointClass, true, confidence) and
+    characteristic2.hasImplications(endpointClass2, true, confidence2) and
+    confidence > characteristic.mediumConfidence() and
+    confidence2 > characteristic2.mediumConfidence()
+  ) and
+  errorMessage = "Endpoint has high-confidence positive indicators for multiple classes"
+  or
+  // An enpoint's characteristics should not include positive indicators with medium/high confidence for some class and
+  // also include negative indicators with medium/high confidence for this same class.
+  exists(EndpointCharacteristic characteristic2, float confidence2 |
+    characteristic.appliesToEndpoint(endpoint) and
+    characteristic2.appliesToEndpoint(endpoint) and
+    characteristic.hasImplications(endpointClass, true, confidence) and
+    characteristic2.hasImplications(endpointClass, false, confidence2) and
+    confidence > characteristic.mediumConfidence() and
+    confidence2 > characteristic2.mediumConfidence()
+  ) and
+  errorMessage = "Endpoint has high-confidence positive and negative indicators for the same class"
+}
+
+query predicate erroneousConfidences(
+  EndpointCharacteristic characteristic, float confidence, string errorMessage
+) {
+  characteristic.hasImplications(_, _, confidence) and
+  (confidence < 0 or confidence > 1) and
+  errorMessage = "Characteristic has an indicator with confidence outside of [0, 1]"
+}
+
 /**
  * A set of characteristics that a particular endpoint might have. This set of characteristics is used to make decisions
  * about whether to include the endpoint in the training set and with what label, as well as whether to score the

java/ql/experimental/adaptivethreatmodeling/src/ExtractNegativeExamples.ql

@@ -26,11 +26,14 @@ DataFlow::Node getSampleFromSampleRate(float rate) {
 }
 
 from
-  DataFlow::Node sink, EndpointCharacteristics::EndpointCharacteristic characteristic,
+  DataFlow::Node endpoint, EndpointCharacteristics::EndpointCharacteristic characteristic,
   float confidence
 where
-  characteristic.appliesToEndpoint(sink) and
+  characteristic.appliesToEndpoint(endpoint) and
   confidence >= characteristic.highConfidence() and
   characteristic.hasImplications(any(NegativeType negative), true, confidence) and
-  sink = getSampleFromSampleRate(0.01)
-select sink, "Non-sink of type " + characteristic + " with confidence " + confidence.toString()
+  // Exclude endpoints that have contradictory endpoint characteristics, because we only want examples we're highly
+  // certain about in the prompt.
+  not EndpointCharacteristics::erroneousEndpoints(endpoint, _, _, _, _) and
+  endpoint = getSampleFromSampleRate(0.01)
+select endpoint, "Non-sink of type " + characteristic + " with confidence " + confidence.toString()

java/ql/experimental/adaptivethreatmodeling/src/ExtractPositiveExamples.ql

@@ -21,5 +21,8 @@ from
 where
   characteristic.appliesToEndpoint(sink) and
   confidence >= characteristic.maximalConfidence() and
-  characteristic.hasImplications(config.getASinkEndpointType(), true, confidence)
+  characteristic.hasImplications(config.getASinkEndpointType(), true, confidence) and
+  // Exclude sinks that have contradictory endpoint characteristics, because we only want examples we're highly certain
+  // about in the prompt.
+  not EndpointCharacteristics::erroneousEndpoints(sink, _, _, _, _)
 select sink, characteristic.toString()