| # | Beat | Timecode | Motion | Logo | Audio | VO / on-screen |
|---|---|---|---|---|---|---|
| 1 | open | 0–7s | spatial-parallax | icon·liquid-chrome | ♪ — | Some problems do not need a chatbot. They need a model that learned your patterns and predicts what comes next. on-screen: Prediction, not chat |
| 2 | legacy | 7.0–14.0s | crossfade-8f | icon·white-knockout | ♪ — | The old way was forecasting on gut and a trailing average, right until the pattern quietly changed. on-screen: The gut-feel forecast |
| 3 | teach | 14.0–21.9s | kinetic-build | icon·liquid-chrome | ♪ — | First we check if you even need training, because a rule or a prompt often beats a model you have to maintain. on-screen: Step 1: is it even ML? |
| 4 | teach | 21.9–28.9s | kinetic-build | icon·white-knockout | ♪ — | A model learns from examples, so the real work is assembling clean, labeled history of what actually happened. on-screen: Step 2: label the data |
| 5 | teach | 28.9–35.9s | kinetic-build | icon·liquid-chrome | ♪ — | We hide a slice of data from training, so we can measure the model on examples it has never seen. on-screen: Step 3: hold out a test set |
| 6 | teach | 35.9–42.9s | kinetic-build | icon·white-knockout | ♪ — | We always build the dumbest possible predictor first, so we know whether the fancy model is actually earning its keep. on-screen: Step 4: start with a baseline |
| 7 | teach | 42.9–49.699999999999996s | kinetic-build | icon·liquid-chrome | ♪ — | We choose the score that matches the cost of being wrong, because accuracy lies when the classes are lopsided. on-screen: Step 5: pick the right metric |
| 8 | teach | 49.7–56.800000000000004s | kinetic-build | icon·white-knockout | ♪ — | We hunt for data that secretly tells the answer, because a model that cheats in testing fails in the on-screen: Step 6: watch for leakage |
| 9 | teach | 56.8–63.599999999999994s | kinetic-build | icon·liquid-chrome | ♪ — | We launch it in shadow mode first, predicting alongside the old way, before we let it touch a decision. on-screen: Step 7: ship behind a gate |
| 10 | teach | 63.6–70.4s | kinetic-build | icon·white-knockout | ♪ — | We monitor it after launch, because the world moves and a model trained on last year slowly goes stale. on-screen: Step 8: watch it drift |
| 11 | teach | 70.4–76.4s | kinetic-build | icon·liquid-chrome | ♪ — | When accuracy slips we retrain on fresh data, so the model keeps up instead of quietly rotting. on-screen: Step 9: retrain on a loop |
| 12 | proof | 76.4–83.2s | receipts-count | icon·white-knockout | ♪ — | That is a model built on a baseline, judged on unseen data, shipped in shadow, and watched for drift. on-screen: Learned, tested, watched |
| 13 | futurist | 83.2–90.3s | spatial-parallax | icon·liquid-chrome | ♪ — | A model that learns your patterns lets you act on what is about to happen, not just what already on-screen: What this unlocks |
| 14 | resolve | 90.3–96.3s | coalescence | stacked·liquid-chrome | ♪ swell | Tell us what you wish you could see coming. We will tell you if a model can. temerarii.xyz. on-screen: Name the prediction |