Part 01 / Software & machine learning
04 / 09Answer B, built
Ask a latent to predict what happens next.
Train one history encoder under four objectives. Test whether prediction forces the latent to represent useful dynamics, without turning a proposed answer into a prewritten result.
The goal for this build
What you’ll be able to do.
- Build a causal 50-step history representation.
- Train a residual prediction objective ten steps (0.1 s) ahead.
- Compare reconstruction, contrastive and compression objectives fairly.
- Probe held-out dynamics without leaking flight identity or labels.
What to have ready
- Modules 1–3 and the fixed benchmark
- Gradient descent and train/validation/test splits
Free account
Read the lesson, run the lab, pass the gate.
The whole course is free. Create an account to unlock the lesson, the build lab and the gate command for this module, and to download the starter kit.
03 / Prove it
Keep the evidence.
Then run the gate.
What you ship
- Causal history encoder and residual forward model
- Reproducible dataset manifest with group-level splits
- All four objective checkpoints and training configurations
- Held-out probe notebook, ablation table and Module 04 evidence
Module 04 gateEvidence required
- Objective B reaches held-out R² above 0.9 for the thrust-to-mass ratio kf/m, per history window, as the median over the same three or more training seeds used for every objective.
- At least 20 held-out vehicles, each at least 2% from every training pair.
- A, B, C and D results are all reported with unchanged splits and comparable budgets.
- No future observations or simulator parameter labels enter the deployment encoder.
With a free account: the gate command and how to record your evidence.
The story behind the build
Episode companions.
Written companions are available now. Recordings and customer access are forthcoming.
Made it through the lesson?
Save your reading progress here. Record your gate result in your project.