Use this guide for the full 120-minute Session 3 run. Treat it as a re-entry and synthesis session, not an advanced graduation test: participants may have attended both earlier sessions, one, or neither. It turns the first two meetings into one argument, tests its video version through comparison, protects one finished participant claim, and hands the final half hour to guest Dr. Emily Thomforde.
Camp journey. Participants have been the machine three times.
Session 1, text. Fluent output is not necessarily true.
Session 2, images. Plausible output is not necessarily neutral.
Session 3, video. Smooth output is not evidence.
Core question. What changes when generation has to preserve relationships across time?
A tracker estimates correspondence across frames that already exist. A generator produces new frames whose subjects, objects, setting, camera, and motion must remain mutually coherent. Tracking is a useful contrast, not a literal description of one universal generation pipeline.
Before the session
- Arrange two participant artifact shares in advance. Ask each person to show one artifact, make one claim, and name one boundary in about three minutes. Include at least one classroom-facing share, and do not select only from the people most visible in chat.
- Send Dr. Emily Thomforde the Session 1 and Session 2 recaps, the three-line argument, the remaining-question list, the 18-minute talk / 8-minute Q&A timing, and the private time signal. Brief her that the likely room includes people who are still figuring the tools out, that bias and learned defaults have been a strong live interest, and that many registrants want classroom transfer. Ask for one beginner-readable example and one practical implication for educators.
- Prepare one anonymous poll or private-writing prompt for the opening. Make chat an option, not the only visible participation channel.
- Test the live deck, Point Correspondence Lab, and Coherence Animator at the same browser and Zoom share size you will use live.
- Prepare two short frozen failure examples: one identity or object-consistency break and one physics, text, or camera-continuity break. Do not rely on live generation.
- Put the Video Test Report and default / advanced studio links in one paste-ready chat message.
- Confirm the guest’s preferred introduction, name pronunciation, and recording consent.
Materials
- Session 3 live deck
- Point Correspondence Lab
- Coherence Animator — core individual activity
- Temporal Telephone — optional facilitator-led or paper fallback
- Video Failure Gallery Viewer
- Frame-by-Frame Coherence Viewer
- Video Test Report
- AI Use + Consent Checklist
- Optional reference: CoTracker for point correspondence across existing frames
Run of show
| Time | Segment | Facilitator move | Participant action |
|---|---|---|---|
| 0–5 | Welcome + re-entry | Repeat the recording / likeness boundary and camera-off, chat-only, private-writing, or watch-only choices. Put the three-line mechanism map and argument on screen. | Respond first through an anonymous poll or private writing. Chat is optional. |
| 5–12 | Two artifact shares | Invite the two pre-arranged participants, including one classroom-facing share: artifact, claim, boundary. | Listen for the evidence. |
| 12–17 | Active synthesis | Retrieve, do not re-lecture. Give 30 seconds of private writing, take one poll or precise chat response, then invite two voluntary spoken responses. | Match one prior activity to one line using any offered channel. |
| 17–24 | Point correspondence | Predict and reveal in Point Lab; distinguish tracking existing frames from generating new ones. | Name one relationship that must persist. |
| 24–47 | Coherence Animator · solo A/B comparison | Demo the interface, then let each person complete one shared opening and four new drawings per run. Run A shows the previous frame; Run B shows the shared opening anchor plus the previous frame. Use guided A twice → B twice playback. | Complete the nine-drawing comparison, track one feature, and name one observation plus one analogy limit. |
| 47–57 | Curated failure hunt | Use two frozen examples and the four-question protocol: what changed, what supports it, what would verify the source, and what can we not conclude? | Name the first exact break and separate observation, explanation, verification, and boundary. |
| 57–65 | System map | Map labor, decisions, consent, and provenance using the same four-question frame. | Name one human decision, hidden contribution, consent question, and provenance question. |
| 65–83 | One-tool studio | Show the worked example, then offer the default and advanced lanes. Require one worksheet row before a second tool. | Finish one observation, bounded claim, missing source evidence, and classroom-use sentence. |
| 83–88 | Two claims | Invite two concise shares selected during the studio, not only through chat visibility. Include one classroom adaptation if available. | Share or listen for observation versus explanation. |
| 88–90 | Guest introduction | Introduce Dr. Emily Thomforde and the listening question. | Close tool tabs and listen. |
| 90–108 | Guest talk | Yield the floor; send the private three-minute signal at 105. | Capture one distinction or question. |
| 108–116 | Q&A | Prioritize participant questions. | Ask concise questions. |
| 116–118 | Guest takeaway | Ask Emily for one sentence to carry forward; thank her. | Put one phrase in chat. |
| 118–120 | Close | Return to the three lines and invite the optional Studio / Showcase. | Save one next step. |
Coherence Animator protocol
- Shared opening. Use the motion prompt to draw and lock Frame 1 once. The exact same image begins both runs.
- Run A — previous only. Draw Frames 2–5 using the previous-frame ghost. The tool stops at exactly five frames.
- Run B — anchor + previous. Draw Frames 2–5 while the shared opening and previous drawing remain visible.
- Choose one feature. Name a hand, face, object edge, position, scale, or camera relationship from the shared opening.
- Compare. Use Play A twice, then B twice at one speed, then toggle A → B → A. Record the first exact change and frame number.
- State the limit. This is a human drawing comparison with one changed visibility condition. It does not reproduce a universal video-model architecture.
Mechanism language to keep
- Say: “Small inconsistencies can compound when a sequence does not preserve the same relationships across time.”
- Do not say: “Drift is the model forgetting.” If you use forgetting as a cartoon, label it.
- Say: “Motion learned from patterns does not guarantee physics.”
- Do not say: “Generated motion is interpolation, not physics.” Some workflows interpolate; the category is broader.
- Observation can support a failure claim. Causal explanation requires model- and workflow-specific evidence.
Framing rules for live discussion
- Anchor mode is not locked noise. The activity tests one simplified control: a stable visual reference. Locked noise, reference conditioning, masks, cross-attention controls, and compositing are distinct ways that workflows may add control.
- Glitches are not authentication. Close observation can find a failure; it cannot verify a clip. Verification needs provenance, source context, corroboration, and, where available, content credentials.
- Bound the sustainability claim. Video often introduces much more spatial-temporal data and substantial compute, but costs vary by architecture, workflow, duration, resolution, and hardware.
- Do not assign intent to the generator. Ask what inputs the system receives, how the interface represents prompt / reference / mask controls, which inputs affect generation, and what the surrounding platform adds.
Studio lanes
Worked example. At frame 4, the cup handle disappears. Comparing frames 3 and 4 supports that observation, but it does not prove what caused the change. Before treating the clip as evidence, I would need its source and corroborating context. With learners, I would pause between the two frames and ask them to separate observation from explanation.
Default lane. Use one frozen example in the Video Failure Gallery Viewer. Identify the first exact break, capture the frame or timestamp, complete the bounded claim, name one piece of source evidence still needed, and add one sentence beginning “With learners, I would…”. This route requires no generation and no second tool.
Advanced lane. Use Coherence Animator, Point Correspondence Lab, or an accessible video tool. Change one reference, camera, duration, motion, or subject constraint; compare the result and state what the comparison cannot prove.
No-AI lane. Analyze frozen frames or draw the same one-opening/two-branch sequence on paper. Temporal Telephone remains a facilitator-led group variation. No account or generation is required.
Four-question verification handoff
- What exactly changed?
- What evidence supports that observation?
- What would verify the clip’s source?
- What can we not conclude?
Use the same questions in the failure hunt, system map, studio, and final claims. A glitch is not proof that a clip is synthetic, and a clean clip is not proof that it is authentic.
What success looks like
- Participants complete bounded claims, including through the default or watch-and-write lane.
- Participants distinguish tracking evidence in existing frames from generating new frames.
- Participants name at least one missing piece of provenance evidence.
- People who are still figuring the tools out can finish the default lane.
- Participation is visible across private writing, polls, drawing / tools, chat, and speech.
- Chat volume is not used as a proxy for understanding or engagement.
Facilitator prompts
- “What exact frame or comparison is your evidence?”
- “What changed, what stayed constant, and what can this test not prove?”
- “What did the anchor preserve? What did it fail to preserve?”
- “What would you need to learn before claiming that a real generator works like this activity?”
- “Is this an observation about the output, or an explanation of the mechanism?”
- “What makes the clip convincing? What would make it trustworthy evidence?”
- “Whose labor, likeness, or data is hidden in the output?”
- “Would a provenance record establish consent, or only document a chain?”
Fallbacks and cut rules
- If shares are not confirmed, use two prepared recap artifacts; do not cold-call.
- If Point Lab fails, use the deck’s fallback ball strip.
- If Coherence Animator fails, use the same one-opening/two-branch sequence on paper, or run Temporal Telephone as a facilitator-led relay. Preserve the Run A / Run B comparison.
- If time slips before the guest, shorten the failure hunt to one example and the system map to four minutes. Do not take time from the guest.
- If the guest has a technical problem, extend the studio and use the prepared Q&A prompts; preserve the final callback.
- If participants finish early, strengthen the boundary of the claim instead of opening another tool.
After the session
- Save the deck log and Zoom chat with the recording, captions, and transcript.
- Capture the two claims as source notes, but do not publish names, quotes, screenshots, or work without the required permission.
- Send the Session 3 feedback link with the Week 4 / Optional Studio commitment, participation-mode, follow-up-format, and “Smooth video is not evidence because…” questions.
- Record what failed operationally, especially guest timing, tool handoffs, whether participants finished one claim, and whether the default lane worked for people with uneven prior experience.