Research on identifiable computational units for AI systems.
AWARE introduces caps (capability nodes): identifiable, lifecycle-managed computational units. Each cap carries a stable identifier across training, supports lifecycle operations (discovery, freezing, replacement, growth), and exposes its learned content for inspection. The cap primitive is general; the current empirical work integrates it into transformer-style language models.
On TinyStories at matched parameter count (~895K), the cap-input architecture with 3-token windows achieves val perplexity 13.71 ± 0.33 versus a pure-transformer baseline of 28.00 ± 0.11 - a 51% perplexity reduction, multi-seed validated.
The paper deliberately covers a narrow, well-defined contribution: the cap primitive and one validated configuration. Continual learning, audit-based replacement, concept stability, and cap-native deep architectures are sketched as integration paths but left for follow-up work.
Beyond the current paper, AWARE explores a longer-term question: what properties does a system need to have to be truly autonomous - self-contained, self-modeling, self-directed? The questions below frame an open research agenda. They are not currently validated; the cap primitive is the substrate on which we plan to build toward them.
Can a system's internal processes produce and maintain every component - including the boundary itself - without external orchestration?
Can a system contain a compressed representation of itself, used to guide its own growth and modification? The recursion would terminate because the model is lossy.
Can goals arise from the interaction between a self-model and a knowledge substrate, rather than being externally specified as a loss function or reward signal?
What does a knowledge structure look like that persists across time, is readable and writable by the system's own processes, and is reorganizable - not append-only, not immutable?
Can a system modify its own parameters in response to experience, in real-time, without separate train and inference phases?
Can a system monitor its own computational substrate (memory, latency, parameter norms, substrate density) and condition decisions on that physical state?
Can no component of a system remain permanently derivable from an external specification? The seed is raw material to be metabolized, not permanent instruction.
Can a system choose its own inputs - expressing internal motivation as outward requests to the environment? Without this, even a system that satisfies the prior seven would remain a passive learner.
Early exploratory essays on AWARE's long-term direction. Written before the current empirical work; the ideas remain open research questions.