A known starting point
E000 reproduced a supplied, fully observed affine-dynamics baseline and received a separate review. Its learner received simulator state directly.
EXPERIMENTS / LEARNED MODELS
World Model Lab
What can a learned model predict, and what does a successful experiment actually establish?
A series of small experiments, each with a specific question and a limit on what its answer can tell us. A world model is an internal representation that can help an agent predict how a system may change, including what its actions might cause. The larger question is whether an agent can learn such a model from observation, use it to plan, and adapt when conditions change.
Start with a bounded question.
The first experiment reproduced a supplied model of a simple system’s changing state. The proposed next experiment asks the learner to work from images. That changes the problem: position and motion would have to be inferred from what is visible.
A different input
E001 proposes using two visual frames and an action. Simulator state would be reserved for evaluation rather than supplied to the learner.
An open result
The visual experiment has not run. Source preparation and review do not establish that a model can learn from the proposed observations.
What the work has shown.
The established result is narrow: the specified baseline was reproduced. E001 asks a harder, still untested question about learning useful predictive state from observations.
A reproduction has a scope
Success with direct access to simulator state says nothing yet about learning a representation from images. The next question needs its own experiment.
Prediction with actions
E000 was given the relevant state. E001 would have to infer it from images, then predict a possible future situation from the current situation and an action. That is a step toward asking whether predictions could support planning, not a planning result.
Where the experiments are going
Observe → predict → plan → probe → adapt → transfer. These are future research questions, not established capabilities. The work is near the beginning of this sequence.
Observe
Can useful predictive state be inferred from observations instead of supplied directly?
Predict
Can it predict how the system changes under actions?
Plan
Can predictions help choose actions toward a goal?
Probe
Can an action reveal something uncertain about the environment?
Adapt
Can an agent notice when its model is wrong and adjust to unfamiliar conditions?
Transfer
Does what was learned help beyond the training conditions or task?
Success at one stage would not establish the others. Simple baselines and negative results can help show where a learned model is useful.
What remains open
No E001 environment installation, dataset generation, training, or experiment was performed in the work described here. The evidence does not establish visual representation learning, autonomous scientific discovery, or general intelligence.
Process note: source review kept generating more work without an effective stopping rule. The field note examines that cycle and the decision to stop.
What did this actually establish?