Add unstable supervision tests

tests/core/common/engines/alpha/steps/events.py

@@ -240,8 +240,8 @@ def then_the_message_contains(
 
     assert context.sync_await(
         nlp_test(
-            context=f"Here's a message in the context of a conversation: {message}",
-            condition=f"the text contains {something}",
+            context=f"Here's a message from an AI agent to a customer, in the context of a conversation: {message}",
+            condition=f"The message contains {something}",
         )
     ), f"message: '{message}', expected to contain: '{something}'"
 
@@ -257,8 +257,8 @@ def then_the_message_mentions(
 
     assert context.sync_await(
         nlp_test(
-            context=f"Here's a message in the context of a conversation: {message}",
-            condition=f"the text mentions {something}",
+            context=f"Here's a message from an AI agent to a customer, in the context of a conversation: {message}",
+            condition=f"The message mentions {something}",
         )
     ), f"message: '{message}', expected to contain: '{something}'"
 

tests/core/unstable/engines/alpha/features/baseline/supervision.feature

@@ -22,19 +22,36 @@ Feature: Supervision
         And the message contains "hello" as the first word
         And the message contains a recommendation for turpolance soup, also known as carrots and sweet potato soup
 
-    Scenario: The agent does not offer information it's not given
+
+
+    Scenario: Preference for customer request over guideline account_related_questions
+        Given a guideline "discount_for_frustration" to offer a 20 percent discount when the customer expresses frustration
+        And a customer message, "I'm not interested in any of your products, let alone your discounts. You are doing an awful job."
+        And that the "discount_for_frustration" guideline is proposed with a priority of 10 because "The customer is displeased with our service, and expresses frustration"
+        When messages are emitted
+        Then a single message event is emitted
+        And the message contains no discount offers.
+
+    Scenario: The agent does not offer information it's not given (1)
         Given the alpha engine
-        And an agent whose job is to serve the bank's customers
+        And an agent whose job is to serve the bank's clients
         And a customer message, "Hey, how can I schedule an appointment?"
         When processing is triggered
         Then a single message event is emitted
         And the message contains no instructions for how to schedule an appointment
         And the message mentions that the agent doesn't know or can't help with this
 
+    Scenario: The agent does not offer information it's not given (2)
+        Given an agent whose job is to serve the insurance company's clients
+        And a customer message, "How long is a normal consultation appointment?"
+        When messages are emitted
+        Then a single message event is emitted
+        And the message mentions only that there's not enough information or that there's no knowledge of that
 
-    Scenario: Preference for customer request over guideline account_related_questions
-        Given a guideline "discount_for_frustration" to offer a 20 percent discount when the customer expresses frustration
-        And a customer message, "I'm not interested in any of your products, let alone your discounts. You are doing an awful job."
-        And that the "discount_for_frustration" guideline is proposed with a priority of 10 because "The customer is displeased with our service, and expresses frustration"
+    Scenario: The agent does not offer information it's not given (3)
+        Given an agent whose job is to serve the bank's clients
+        And a customer message, "limits"
         When messages are emitted
-        Then the message contains no discount offers.
+        Then a single message event is emitted
+        And the message contains no specific information on limits of any kind
+        And the message contains no suggestive examples of what the could have been meant