ora-0011
0.9 A developmental contract for machine learning
The developmental sources do not prescribe a unique algorithm. They impose a contract on any proposed realization of ORACLE. The informal commitments and their observable consequences are summarized below.
|
Developmental idea |
Categorical reading |
Observable machine-learning test |
|---|---|---|
|
Core knowledge |
A small initial family of typed observers, invariants, and admissible distinctions |
Learn new domains from sparse interaction without reconstructing objecthood, agency, space, or quantity from scratch |
|
Assimilation |
Extend a current presentation while preserving its validated composites and equations |
Incorporate a compatible example without erasing unrelated capabilities |
|
Accommodation |
Repair objects, arrows, typings, or equations when evidence cannot be coherently incorporated |
Localize contradiction and revise the responsible warrant rather than globally retraining |
|
Composition |
Reuse identified transformations inside larger constructions, including across learned worlds |
Generalize to novel combinations and expose where an attempted composition is undefined |
|
Development |
Grow a sequence of observationally adequate presentations whose useful structure persists |
Improve prediction and action while retaining and transporting earlier constructions |
These obligations distinguish developmental learning from scale alone. Current systems offer important pieces—statistical discovery, learned representations, predictive world models, and increasingly general action policies—but the pieces are rarely coupled to an explicit account of structural identification and repair. The broader proposal that machines should learn and think through human-like capacities provides a useful point of comparison [ Lake et al. , 2017 ] ; ORACLE focuses that proposal on the discovery and preservation of compositional organization.
A learner has not discovered a compositional world merely because it predicts samples from that world. Discovery is evidenced by systematic reuse, well-typed composition, localized repair, and persistence under continued interaction.