FROM A PROJECT TO EVIDENCE

Make a claim.
Design its test.

Turn the ebook’s twelve-week implementation plan into a reproducible experiment. Start with exact environments and strong baselines. Add learning only when you can measure what it changes.

Optional research extension. Twelve weeks is the source’s suggested research schedule, not a required course duration. The starter below implements two environments and non-neural baselines; later stages describe work to build and test yourself.

Follow the implementation ladder

Weeks 1–2

Define environments

Build: Typed state/action contracts for two or three games; exact move validators; reproducible starts; separate development and reserved sets.

Ask: Which observations distinguish states? Can the validator reject illegal moves independently of a policy?

Evidence to collect: Run contract and transition tests. Save suite version, initial states, rules, and split membership before tuning.

Available in this course: The starter implements grid navigation and a 2×3 sliding puzzle with exact transitions and deterministic task starts.

Your implementation step: Add a third domain only after its rules and terminal conditions are independently testable.

Review typed interfacesEvidence and questions for Weeks 1–2

These are your notes, not verified completion records. Say “not run” where evidence is missing.

Inspect a shared action contract

Both environments accept action indices 0–3 for up, right, down, and left. In the grid, the actor moves; in the sliding puzzle, the blank moves. The state’s family, rule, layout, and values identify the task. A shared schema does not mean shared learned knowledge.

EnvironmentInspection partitionCase

Grid: actor at cell 22, goal cell 24; blocked cells 2, 7, 12, 17. Cells count from zero, left to right, top to bottom.

Current typed task

{
  "family": "grid",
  "rule": "standard",
  "layout": 1,
  "state": [
    22
  ]
}

Identity: grid/standard/1/22. transition(task, action) returns a new state, or null for an illegal move; it does not mutate its input.

Predict which direction is illegal before trying it. Grid “×” cells are blocked; “G” is the goal. Sliding tile 1 can only move vertically under the changed rule. Grid changed-context cases use a different wall layout.

Compare before adding a model

Each partition has eight grid tasks and eight sliding tasks. Standard training, development, and reserved starts are disjoint by full task identity. Paths may overlap. Changed contexts may reuse state coordinates under different layouts or rules. This is not an evaluation on entirely new game families.

The exact goal-reachable task generator is inspectable in the downloaded source. Reserved cases are protected by workflow here, not encrypted or secret. Inspecting source cases while tuning counts as exposure.

Expansion budget per case

All four methods cap paths at 32 moves. Each expanded node checks four directions; execution checks are separate. Equal expansion limits do not equalize total compute: heuristic arithmetic, queue management, and other overhead are excluded from query counts. Greedy memory prevents revisits but does not add backtracking. BFS without memory can waste its budget revisiting states.

Write the protocol before opening reserved cases

Specify a question, the evaluation split, costs, decision criteria, and limits. The example is editable and describes this small memory-ablation study. It is a starting point, not a validated research proposal.

Research questionName a baseline, a change, and a falsifiable prediction.Write at least 20 characters · Length does not assess research quality.Data and noveltyName training, development, reserved, and changed-context cases. Explain what can overlap.Write at least 20 characters · Length does not assess research quality.Resource protocolSpecify limits and measured costs, including work excluded from the counters.Write at least 20 characters · Length does not assess research quality.Decision criteriaDefine what would count as useful improvement and unacceptable subgroup regression before seeing reserved results.Write at least 20 characters · Length does not assess research quality.Limits and follow-upState what this sample cannot establish and what new evidence the next claim needs.Write at least 20 characters · Length does not assess research quality.

Freezing locks the expansion budget and protocol. Stage notes remain editable so you can record outcomes. Resuming development keeps known reserved exposure. Comparison results are temporary; export them before leaving this page.

Take a reproducible starting point

The starter is dependency-free JavaScript with typed JSDoc contracts, exact transitions, baseline solvers, deterministic splits, and self-checks. It implements the first two research stages. The later neural and external-controller experiments require your own implementation; existing lessons supply smaller demonstrations.

Editable protocol and stage notes travel with My Work progress backups. This extension never changes required course completion.

Source: ebook appendix C, pages 81–82. Return to the architecture lesson or agent project.