Test whether the model helps action.

This is the same reference used in Technical mode. Expand it whenever you need the exact numbers, assumptions or further explanation.

Give the model a job

A useful world model should improve a declared task compared with a relevant alternative. Predicting the next position, preserving a long trajectory and reaching a target are related but distinct. The lab reports them separately instead of combining them into a reassuring single score.

The model fits only the original 96 training transitions. Freeze the collection coverage and planner horizon before comparing outcomes. A later control edit does not relabel an existing result. The fitted affine model and planner receive observed current state, not future simulator states.

Three protocols, three meanings

One-step position error compares a prediction from the true current input with the next independently computed position. Mean absolute error has units of metres. A twelve-step free rollout receives only its initial state and the declared +1/−1 N action schedule; final position discrepancy measures accumulated prediction error.

Control runs twenty exact simulator intervals. Receding-horizon beam search proposes an action using its fitted model; the external evaluator executes that action. A fixed-gain feedback controller supplies a serious comparator. The task succeeds only after all twenty intervals with absolute position below 0.2 m and absolute speed below 0.3 m/s. Low cost, a plausible prediction or an unfinished trace never substitutes for that success rule.

The cost uses squared position, 0.1 times squared speed and a small normalized force penalty. It expresses the stated objective, not physical safety. Model-call totals expose additional planning work. A larger horizon can improve cost while still failing the task threshold or spending substantially more computation.

Change the world without changing the evaluator

Two starts each run nominal and changed dynamics. Nominal mass is 1 kg with linear drag 0.2; changed mass is 1.4 kg with linear drag 0.35. Both retain quadratic drag 0.6 and the same velocity-first interval rule. The model is not refitted on changed outcomes.

The independent evaluator computes $v'=v+0.1(u/m-dv-0.6v|v|)$ and $x'=x+0.1v'$. Here m is mass and d is linear drag; primes denote the next interval. For example, heavier mass changes how much a fixed force changes speed. The planner’s internal prediction cannot alter this external contract.

Freeze, expose and report honestly

Use development comparisons to choose a recipe. Opening reserved starts supplies a separate comparison, but once viewed they are exposed. Public client-side source is inspectable; reserved is a protocol distinction, not secrecy protection. The temporary lesson remembers exposure only during this visit. The saved project retains known exposure and the first reserved recipe through reload and import.

Keep every failed case and distinguish a report with complete fields from independently assessed reasoning. The cumulative project adds separate numerical transfer cases and an implementation bridge. A supported decision to reject model-based planning is a valid conclusion when the baseline is preferable under the declared conditions.