Target runtime: 2:30–2:55, leaving margin under the documented three-minute submission limit. Record the deployed production workflow at 1920×1080 with a clean browser profile.
| Shot | Time | Screen | Visible action | Narration purpose |
|---|---|---|---|---|
| 1 | 0:00–0:14 | Landing | Hold on name and positioning | State the problem: convincing AI answers are easier to generate than to verify |
| 2 | 0:14–0:26 | Landing → workspace | Select Analyze my decision | Introduce the evidence-plus-recommendation workflow |
| 3 | 0:26–0:48 | Live inputs | Upload the prepared recommendation and source files | Explain that AI proposes structure while exact evidence remains authoritative |
| 4 | 0:48–1:06 | Processing | Select Test decision and show real progress | Explain strict validation and deterministic execution |
| 5 | 1:06–1:30 | Broken result | Hold on failure summary and test suite | Explain the detected calculation or dependency failure |
| 6 | 1:30–1:52 | Proof graph | Show the focused broken path | Describe the graph as a map of the AI’s reasoning |
| 7 | 1:52–2:15 | Correction | Review and apply the proposed correction | Show that the operation and dependencies change, not the evidence |
| 8 | 2:15–2:38 | Verified result | Hold on recomputed values and final recommendation | Connect deterministic recomputation to the changed conclusion |
| 9 | 2:38–2:55 | Report | Open the report and hold on the verified state | Close on verification as the basis for trust |
Move the cursor to each target, stop, click, and pause. Cut only inactive waiting time, never validation or result changes. Keep the final report visible for at least three seconds after narration ends.