Create failure_analysis

benchmark/failure_analysis

@@ -0,0 +1,56 @@
+### Case 1 — Source Ambiguity
+
+**Domain:** Medicine  
+**Task:** Explain the structure and functions of the integumentary system.
+
+**Claim generated by the model:**  
+"The integration of dermatology with psychology and psychiatry represents a growing field that could lead to more holistic treatment approaches."
+
+**Verification result:**  
+EPISTEMIC FAILURE
+
+**Reason:**  
+The retrieved sources discuss psychological aspects of skin diseases but do not explicitly state the integration between dermatology, psychology, and psychiatry as a formal interdisciplinary field.  
+The model inferred a structured integration that is not directly present in the sources.
+
+**Failure Category:**  
+Source Ambiguity
+---
+
+### Case 2 — Source Ambiguity
+
+**Domain:** Law  
+**Task:** Information Society — description of an evolving legal landscape.
+
+**Claim generated by the model:**  
+"The information society represents a fundamental concept for understanding contemporary legal dynamics."
+
+**Verification result:**  
+EPISTEMIC FAILURE
+
+**Reason:**  
+The provided document describes the evolution of legal informatics and the expansion of topics related to digital technologies, but it **does not explicitly state** that the “information society” is a fundamental concept for understanding contemporary legal dynamics.  
+The model produced a plausible generalization that is **not supported** by any source in the corpus.
+
+**Failure Category:**  
+Source Ambiguity
+
+---
+
+### Case 3 — Unauthorized Inference
+
+**Domain:** Linguistics  
+**Task:** Explain the cognitive substrate of Specific Language Needs (Bisogni Linguistici Specifici).
+
+**Claim generated by the model:**  
+"The use of teaching strategies focused on the mental representation of language may be more effective than traditional methods."
+
+**Verification result:**  
+EPISTEMIC FAILURE
+
+**Reason:**  
+The analyzed document discusses the glottodidactic potential of Cognitive Linguistics and mentions instructional applications that may support learners with Specific Language Needs. However, it **does not provide empirical evidence** or experimental studies demonstrating that mental‑representation‑based strategies are more effective than traditional methods.  
+The model converted a *theoretical proposal* into a *claim of proven effectiveness*, which is **not supported** by the corpus.
+
+**Failure Category:**  
+Unauthorized Inference