Answer B, built

Make a latent predict what the arm does next.

Train one history encoder on four objectives and let a linear probe judge them. Objective B predicts how the real arm departs from the nominal model 0.1 s ahead, and its latent should reveal m/kₜ, the payload mass relative to the torque constant.

Written lesson + build labEvidence-based gate

The goal for this build

What you’ll be able to do.

Build a causal encoder over the last 50 policy steps (0.5 s) of observations and goals.
Train objective B to predict the residual between the real state 0.1 s ahead and the nominal model’s rollout.
Train objectives A, C and D on the same encoder, data and budget, and report all four.
Probe m/kₜ on held-out arms, and explain why currents cannot reveal the mass alone.

What to have ready

Module 1 sim with a Params dataclass you can randomize
Module 3 EKF, and why a static hold cannot tell payload mass from torque constant
Gradient descent, and train, validation and test splits
JAX with Flax and Optax running on your laptop CPU

Download the course starter ↓

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, the task and the gate command for this module, and to download the starter kit.

Create a free account Sign in

03 / Prove it

Keep the evidence.
Then run the gate.

What you ship

A run generator and a split manifest of whole arms, with parameters, seeds and config hashes
The causal history encoder and residual forward model for objective B
The same encoder trained on A, C and D, with three or more seeds per objective
A probing notebook that regresses true parameters from z, with per-window held-out scores
Ten sim trials of Which cup is heavier?, and the module’s PROGRESS.md line

Module 04 gateEvidence required

With objective B, a linear probe recovers the ratio of payload mass to torque constant, m/kₜ, with R² above 0.9 on held-out arms.
The score is taken per history window, as the median over the same three or more training seeds for every objective.
A, C and D are reported side by side with B, whatever they show.
The evidence is the probing notebook’s per-window held-out predictions for every objective and seed, each with its commit, config and seed.

With a free account: the gate command and how to record your evidence.

The story behind the build

Episode companions.

EP 04A, B, C or D? I trained all four EP 05Building the history encoder

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.