Audience Guide

• Web Version (PDF) please click here to view.

• Audio (5 min) click here for audio file OR call 249-700-8494 and press 1

Performance

You can submit your prompts via Web or SMS before and during show. They will only become available about 10 minutes before the showtime:

• Web please click here to access a private chat. <<— CLICK HERE TO SUBMIT IDEAS TO WEB CHAT
  

• SMS - text 249-700-8494
  When texting, to prevent spam you must sent the SHOWCODE with each message.
  
  Example for if the SHOWCODE is “MM”, text “MM help” to request AI generated assistance or enter “MM banana” to suggest banana.

Post-Show Feedback & Support

If you would like to share feedback about your experience or suggestions for future performances, please fill out the form here or please reach out directly to Amanda.

Given the unpredictable nature of this show, we want to provide you with information about after-care:

• For immediate mental healh assistance, please call or text Canadian Mental Health Crisis Line 9-8-8 or seek emergency support.

• For non-urgent requests, please contact Amanda or the Fringe Festival Staff for additional resources and after-care support.

Additional Technical Notes

• Show Access: Each performance has a unique, time-limited access code that allows audience participation. For SMS, all messages must contact showcode as a prefix.

• Performance Duration: Audience will have access to submit their prompts approximately 5 minutes before showtime and throughout the show. When the performer begins the show and starts to receive audio clips to the headset, there is a timer which will halt the show at the end of the target runtime.

• Audience Input Pipeline: Audience responses are collected anonymously via both web form and SMS. All incoming data is submitted into a central Redis-backed queue.

• Response Selection Logic: A background worker selects and combines between 1 to 4 audience inputs at random. This aggregation forms the core prompt for the AI.

• Generate Script System Prompt: LLM system prompt is configured to generate a short narrative for a solo performer at an accessible level, to emphasize user input as the primary theme of the of segment, keep the scenarios fun, heartfelt, and intriguing as well as intructions to permit sensitive topics for a mature audience. More details available upon request.

• Help Request System Prompt: When audience submit the word “help” it will trigger an LLM prompt to generate ideation assistance. The base system prompt for all Help Requests is as follows: “Top tier customer service helping people develop new ideas by asking {adjective} questions. Provide 2 plain language questions with NO PRETEXT or COMMENTARY OR NUMBERING. Just the questions.” where each request randomly selected from a set of 9 AI generated adjectives such as: “playful”, “puzzling”, “socratic”.

• LLM Model: Text is generated using Google's gemma-2-2b-it language model via Hugging Face Inference APIs

  • Temperature Sampling: Temperature for text generation is drawn from a normal distribution with a target mean of 0.7, standard deviation of 0.25, and is clipped to remain within the range [0.4, 1.2]. This introduces subtle variation while maintaining performance stability.

• Architecture:

  • Backend: Python Flask web service

  • Deployment: Hosted on Render

  • Queue Management: Redis Queue (RQ)

  • Text To Speech: Google Text-To-Speech

  • SMS & Phone Interaction: Twilio

  • Feedback Forms: Google Forms (no emails collected)

• Data Handling:

  • All responses and generated text are stored securely in a private Google Cloud Storage bucket.

  • No user-identifying information is stored; data remains anonymous.

• Post-Show Cleanup: After each performance concludes, all relevant Redis queues are purged to reset the system for the next show and maintain data hygiene.