Impromptu Oral Assessment Blueprint

1. Overview & Pedagogical Philosophy

The Impromptu Oral Bookend model pairs standard objective or applied exam questions with short, time-boxed audio/video responses at the start and end of an online exam.

Rather than relying on invasive, expensive proctoring software (which fails to observe second screens, phones, or off-camera assistance), this model attacks the cognitive pipeline of cheating. It shifts the focus from “catching students looking away” to authenticating the learner through spontaneous spoken reasoning.

+------------------------------------------------------------------------+
|                        MOODLE EXAM WORKFLOW                            |
+------------------------------------------------------------------------+
| 1. IDENTITY & CALIBRATION (Start)                                      |
|    - 60-Second Video Viva (RecordRTC)                                  |
|    - Randomized Physical Gesture (Pen, Mug, Notepad, Clothing anchor)  |
|    - Shatters deepfakes, teleprompters, and establishes voice baseline |
+------------------------------------------------------------------------+
| 2. CORE EXAMINATION BODY (Middle)                                      |
|    - Objective, Scenario, or Problem-Solving Questions                 |
|    - Standard teacher question bank                                    |
+------------------------------------------------------------------------+
| 3. RETROSPECTIVE COGNITIVE DEFENSE (End)                               |
|    - 90-Second Spoken Defense (Auto-submit on timer expiry)            |
|    - Explain reasoning for a chosen question in everyday words         |
|    - Destroys second-device and AI transcription outsourcing           |
+------------------------------------------------------------------------+

2. Why It Defeats Generative AI & Second Devices

A. The Physical Occlusion Anchor (Breaks Video AI & Filters)

When students attempt to use live deepfakes, AI avatars, or digital eye-tracking teleprompters (such as Nvidia Broadcast or OBS virtual cams):

Facial Occlusion Glitch: Prompting the student to hold an everyday object (e.g., a pen or notepad) directly across their chin or cheek forces an occlusion. Most real-time neural face-swappers immediately warp, tear, or reveal the true underlying face.
Teleprompter Eye-Track Disruption: Asking the student to look up at the ceiling or glance at a physical item breaks software designed to keep simulated eyes locked on the camera lens.

B. The Cognitive Time Squeeze (The Chokepoint)

Generating an answer via an LLM on a second device requires a student to:

Type or speak the prompt into the phone/laptop.
Wait for the LLM output.
Read and internalize the response.
Translate it into spoken words without sounding scripted.

In a 90-second hard-capped viva, this sequence is impossible to pull off smoothly. Every second spent consulting an AI eats directly into their submission window.

C. Authentic Imperfection vs Scripted Reading

Academic integrity is verified by how ideas are articulated:

A student reading an AI response speaks in uniform cadences, with rigid sentence syntax, unblinking eyes, and zero natural hesitation.
A human who understands the material stumbles, uses conversational analogies, pauses to gather their thoughts, and links theories back to specific lecture discussions.

3. Physical Anchor Permutations for Question Banks

To prevent students from anticipating the verification action, question banks should randomize across four standard physical anchors:

Option 1: The Pen Anchor

Lock Step: Before answering, pick up a pen or pencil. Hold it horizontally directly in front of your chin for 3 seconds while looking straight into the camera lens, then begin your response.

Option 2: The Notepad Anchor

Lock Step: Before answering, hold up your notepad or a blank piece of paper to cover your face for 1 second, lower it to your desk, make direct eye contact with the camera, and begin speaking.

Option 3: The Mug / Bottle Anchor

Lock Step: Before answering, lift your mug, glass, or water bottle to your mouth as if taking a drink, place it back on your desk, look into the camera lens, and begin your answer.

Option 4: The Clothing Anchor

Lock Step: Before answering, use your right hand to firmly touch your collar or shoulder. Hold your hand there for 3 seconds while looking directly at the camera, then begin your explanation.

4. Automated ExamNode Provisioning

Unit Coordinators do not need to configure these questions manually. When provisioning an exam via the ExamNode API or CLI:

# Provision exam with automated oral bookends
php local/examprovision/cli/provision_exam.php \
  --code=ICT302 \
  --name="Final Exam" \
  --oralbookends \
  --xmlfile=core_questions.xml

ExamNode automatically injects Question 01 (Identity & Physical Check) at the start and Question 99 (Cognitive Defense Viva) at the end.

5. The AI-Resilient Evaluation Rubric

Criteria	Proficient (Full Ownership)	Developing (Scripted or Shallow)	Unacceptable (High Risk / Flagged)
Physical Identity Anchor	Successfully performs the randomized gesture within the first 5 seconds.	Attempted gesture, but cut off by camera frame or delayed.	Completely skipped or ignored the randomized gesture. Triggers manual hold.
Local Context Anchoring	Mentions specific course elements (e.g. “As shown on slide 4 of the lecture…”).	Speaks in generic textbook definitions with zero local course reference.	Fails to link concept to the course framework; uses foreign terminology.
Cognitive Fluency & Delivery	Natural speech flow, conversational phrasing, pauses to think (Authentic Imperfection).	Monotone delivery, formal syntax, eyes tracking lines on a secondary screen.	Long pauses followed by verbatim reading of formatted text; zero natural disfluency.
Spontaneous Application	Addresses core nuance of their chosen question within the 90-second window.	Explains theory correctly but struggles to link it directly to their answer.	Cannot explain why their answer was selected; repeats generic definitions.

As of the latest local_examprovision release, every exam provisioned with --oralbookends automatically pairs the Retrospective Cognitive Defense (Question 99) with a preceding Rough Notes / Outline question (Question 98 — “Pre-Defense Brainstorming and Outline”). No manual question bank setup is required.

How It Works:

The Brainstorm (Question 98, ~60 Seconds): The student is presented with an essay text box and asked to jot down 2-3 bullet points outlining their reasoning for one chosen exam question.
The Defense (Question 99, 90 Seconds): The student advances to the final RecordRTC question and verbally explains their logic, explicitly instructed to use their own notes from the previous page as a guide.

Why This Is Effective:

Reduces Cognitive Load & Anxiety: Students often freeze when staring at a recording light. Providing a structured space to outline their thoughts first results in a more natural, conversational delivery on camera.
Dual-Modal AI Auditing: The Ollama evaluation task (evaluate_attempt) that runs after submission reads both the Question 98 text and the Question 99 transcript/response summary into the same evaluation prompt. A candidate reading an LLM script from a second device will typically leave the notes box empty or paste a perfectly formatted paragraph into a box designed for rough bullet points — both are surfaced as anomalies in the AI Integrity verdict on the teacher review screen.
Keystroke Dynamics (Future): If integrated with a keystroke biometrics tool (e.g. tiny_cursive, per the Cursive PoC under discussion with Joe), the brainstorm box would provide an additional data point to verify authorship against the student’s baseline typing signature. Not yet implemented — see the Pre-Flight Audit guide for current automated checks.

Provisioning

No extra flag is needed — Question 98 is injected automatically whenever --oralbookends is set:

php local/examprovision/cli/provision_exam.php \
  --code=ICT302 \
  --name="Final Exam" \
  --oralbookends