Question
Can a learned action-outcome model adapt its means when the state of a supplied task changes?
Experimental setup
A small categorical world supplied action paths, needs, values and a task. The reported comparison covered 1,024 worlds with eight trials per task. The agent learned how actions led to outcomes.
The comparison
Compare consequence-based revaluation against frozen needs and against a complete cache of all 24 policies. The cache tests whether recursive planning is necessary to explain the result.
Result
| Learned consequence model | 98.05% | |
|---|---|---|
| Frozen needs | 33.40% | |
| Complete policy cache | 98.05% |
The learned model achieved 98.05% reported completion against 33.40% for frozen needs. The complete 24-policy outcome cache matched it, including trajectories. This supports consequence-sensitive adaptation of means, not a unique advantage for recursive planning.
What remains open
Goals and values were supplied. The result does not demonstrate acquired ends, genuine desires, phenomenal experience or a live Luna integration. A preceding planning comparison also failed to establish an advantage over a myopic alternative.
Next test
Transfer the smallest useful mechanism into a continuing project and test a delayed later choice. Keep the matched cache or simpler controller wherever it explains the behavior.