Build a searcher and a belief filter.

1. Predict before implementing

This browser preview uses the established TypeScript lesson models; it does not run the Python file or your implementation. The reference CLI uses these same registered teaching settings when producing importable evidence.

2. Download the exercise and public checks

Unzip into a local folder. Use Python 3.10 or later; these files need only its standard library. No package installation, paid API, network service or GPU is required. Save earlier outputs rather than overwriting them.

python3 planner_checks.py planner_reference.py
python3 planner_checks.py planner_exercise.py

The reference passes six tests. The exercise deliberately raises NotImplementedError until you implement both functions. These commands execute local Python normally; they do not use the browser's restricted interpreter.

3. Implement, inspect, repair

1. Implement directed A* with a priority queue. Keep the best known path cost, ignore stale queue entries and reopen a state when a cheaper path is found. Count outgoing edge reads separately from expanded states.
2. Return distinct found, budget and exhausted statuses. Selecting an already-reached goal at the expansion cap requires no further expansion. An overestimating heuristic does not guarantee an optimal route.
3. Implement belief prediction using from/to transition rows. Multiply by current observation likelihoods and normalize. Ones mean no sensor information; zero total weight means undefined conditioning and must stop.
4. Keep inputs unchanged, reject invalid probabilities/negative road costs and enforce the documented work bounds. Add at least one failing example of your own, then repair it without deleting the test.

The supplied checks include 25 reproducible five-node graphs and a different binary hidden-state model. Their oracles enumerate simple paths and latent state sequences rather than calling the learner implementation to generate expected answers. All cases and answers are inspectable public practice; passing them is not a hidden examination or proof of independent authorship.

4. Replay supported evidence in the saved project

python3 planner_reference.py search uniform.json --method uniform
python3 planner_reference.py search misleading.json --method misleading
python3 planner_reference.py belief moving.json --motion slow --steps 3
python3 planner_reference.py belief fixed.json --motion fixed --steps 3

Import the four outputs in the planner project. They are produced by the supplied reference, not your learner exercise. The browser validates supported settings and recomputes the search/filter results; it does not run or authenticate Python code. Keep that provenance explicit.

Generic exercise graphs and binary-filter results do not fit the registered lesson importer. Attach your own source, tests and results as implementation evidence instead. A human reviewer must examine that artifact and test unfamiliar cases. Do not treat imported settings or report submission as proof your implementation is correct.

5. Explain what your evidence establishes

Reference bounds: 32 states, 128 edges and 256 search expansions; 2–8 hidden states and 16 observations. The CLI emits only the registered map and three-state settings. Equal expansion counts do not equal total compute. Model correctness, algorithm correctness and a good real-world decision are separate claims.

Write a new inspect-versus-act example with a prior, both signal branches, an observation cost and the best direct action. Explain search failure, model misspecification and missing evidence separately. The saved project supplies four separate objective transfer cases; runnable-code review and learner pilots remain pending. This optional studio does not add a required completion percentage.