PYTHON · NUMPY

Benign in calm air

The labels carry a measured slope-sixteen systematic error. Clone from them and evaluate in calm air. Region drops by:

Clone the same demonstrations twice, from true and from reconstructed labels, and measure the provably wrong labels costing nothing in calm conditions.

01Challenge

Sixty-four episodes at kp=1.00, two clones, identical everything except the label source. Evaluate both in the training conditions and explain the result.

Predict firstThe labels carry a measured slope-sixteen systematic error. Clone from them and evaluate in calm air. Region drops by:

02Model

The last lesson proved the labels lie with slope sixteen at this stiffness. So clone from them and watch the failure, except there is none. Both clones score a perfect region. The reconstructed-label clone even tracks a hair better. This is the result that keeps the lie invisible at industrial scale, and the mechanism deserves to be exact. What the observer misbooked is the spring's restoring torque, a force that always points back toward equilibrium. A clone that faithfully reproduces those labels applies the restoring force twice: once itself, once via the spring it sits on. In calm air, doubling a restoring force changes the ride's stiffness and changes nothing about where it settles. The misattribution is real, measured, and free. Every evaluation run in the training distribution will bless it. Hold that thought with both hands, because the next lesson changes one number in the environment and nothing in either clone.

03Guided practice

Run both clones. Confirm the gap is zero at region and negligible at tracking, then say precisely why a doubled restoring force forgives.

04Feedback

PASS WHEN You measured provably wrong labels costing zero region and nothing at tracking.

If your run fails, check:

Region 1.000 against 1.000, tracking 0.376 against 0.369. The slope-sixteen lie is behaviourally invisible here, and every training-distribution benchmark would certify this clone.
The misbooked torque is restoring: it points home by construction. Doubling a restoring force retunes the ride and forgives the theft. The lie is only safe while the substrate's work averages zero.
The research bench registered the opposite prediction, that the gap would open with disturbance level, and falsified it: this forgiveness was a finding the bench did not expect. Honest scoring is what made the next lesson's question askable.

05Retrieve & space

Why does doubling a restoring force not change where the system settles?
What single property of the misbooked torque makes the error forgivable here?
Why would every standard evaluation of this clone come back green?

06Mastery & project

State the precise condition under which label misattribution is behaviourally free, and design the evaluation that would stay green while the condition quietly held.

Take an imitation policy you know that passes its evaluations. Write down what its labels could be systematically wrong about that its test distribution never exercises.

← The observer's physics Fatal under load →

PYTHON · NUMPY

Benign in calm air

One demonstration set, two label columns, two clones, forty-eight disjoint test episodes each. The only difference between the clones is who wrote their labels.

You read

the arrays and values already in scope

You change

the code you write in each cell

Fixed

the dataset and the checks that grade you