Optional implementation bridge

Reproduce a model-based policy study

The browser course remains available without installing anything. This dependency-free Python 3.10+ bridge is for learners choosing implementation depth. Download the original starter. It supplies a reference mode and a separate learner TODO; running the reference is guided execution, not independently assessed implementation.

Run and inspect

python world-model-study.py --check
python world-model-study.py --output my-study.json

The output file must be new: existing files are never overwritten. Six whole synthetic episodes provide 72 training transitions, deliberately a different collection from the browser’s 96. The reference fits a ridge affine delta model, optimizes a bounded two-weight policy using actual imagined trajectories, and compares simulator-trained and zero policies. A separate width-six, horizon-three beam MPC reference replans through the frozen learned model before each external step; its 30 model calls per plan are reported separately. Each candidate is frozen before four external evaluation conditions. Output retains every case, parameters, trace and explicit calls.

Implement and falsify

1. Read learner_predict. Implement the affine delta forward pass, independently test the bounded planner’s first-action choice and work count, then run --exercise. Untouched learner mode intentionally fails.
2. Write a separate scalar test for velocity-first motion and check position/velocity units. Derive a known affine update by hand.
3. Add a missing-force coverage intervention to training only. Freeze coefficients and policy before evaluating different starts and changed mass/drag.
4. Record all seeds or deterministic settings, failed runs and work counts. Compare one-step error, rollout fidelity and actual control cost separately.
5. Ask another person to replay your artifact and propose an unfamiliar scenario. Record who reviewed which files and what changed. Until that happens, review stays pending.

Evidence to submit

Attach source and tests, the complete JSON result, a frozen protocol and a supported conclusion to the world-model project. The browser code field stores text inertly; full files remain local. Explain why a model-based gain or loss is plausible, account for collection and optimization costs, and state where the finite case set cannot support a claim. A supported negative result is acceptable. This is a tiny scalar study, not a Dreamer reproduction or physical-robot validation.

Return to imagined policy learning · Return to frozen evaluation