NLU fallback rules getting ignored by rasa

I have created a simple bot where I added recently the fallback classifier. I added it to the end of my pipeline by adding the following to my config.yml file:

  - name: FallbackClassifier
    threshold: 0.3

I also enabled the rules by adding:

policies:
  - name: MemoizationPolicy
  - name: TEDPolicy
    max_history: 5
    epochs: 100
    constrain_similarities: true
  - name: RulePolicy

To always handle the intent by responding with a predefined message I added a simple rule to rules.yml:

# handle fallback and chitchat
- rule: Handle fallback gracefully
  steps:
  - intent: nlu_fallback
  - action: utter_fallback

It seems like it is working fine most of the time, but every now and then the intent “nlu_fallback” is followed up with a different random action. I have double checked and validated that there is no ambiguous state or conflict between rules and/or stories and I am not sure, why rasa would ever decide to use another action than defined in the rule.

Here an example of a conversation that went wrong:

Is there any parameter or option that allows Rasa to ignore rules in certain circumstances? How can I make sure that the intent of “nlu_fallback” is always followed by the action “utter_fallback”?

Many thanks for your help!

I am using rasa 2.6.3 and rasa x 0.40.1.

I think you might be in a form. Does it work if you add nlu_fallback to the ignored_intents of the form?