Implementation status and limitations

Cognia is pre-1.0 research software. This page is part of the contract: it distinguishes stable working paths from experimental mechanisms and syntax that is not fully executed yet.

Working end to end

• parsing, semantic validation and concrete topology expansion;
• populations, nested components and major connection patterns;
• per-neuron leak, bias and gain;
• continuous activation and stochastic firing;
• recurrent connections and per-edge delays;
• fixed, Hebbian/temporal and Oja edge policies;
• chemistry arrays and chemistry-modulated weights/rates;
• workspace competition, ignition, focus, habituation and attention state;
• bounded structural growth configuration;
• signal, text and classification outputs;
• stateless, stateful and raw sequence HTTP execution;
• supervised shallow readout training, including softmax cross-entropy;
• module polarization and controller-driven polarization;
• graph/state save and load for compatible dimensions;
• GPU-backed visualization and introspection.

Experimental or constrained

• Polarization and controller-driven polarization have controlled examples but no broad benchmark or peer-reviewed validation.
• Structural growth is implemented but needs stronger long-run stability and reproducibility studies.
• CHL does not provide reliable deep credit assignment in arbitrary asymmetric graphs.
• Binary network compatibility is dimension- and version-sensitive.
• Runtime layers are inferred from distinct leak values, not connectivity or an explicit layer declaration.
• Sensory routing in the full app is ordinal; source names do not select camera/audio/text routing.
• The HTTP server is a single stateful instance, serializes calls and provides no authentication or TLS.
• The raw HTTP input bypasses cortex preprocessing.
• GPU architecture is configured for the build target and may need rebuilding for another GPU.

Parsed or declared beyond current execution

• Declarative clamp/training phases are not the primary executed training path; command-line trainers implement current reliable training behavior.
• Event declarations and event-driven chemistry/controllers are not all fired end to end at runtime.
• A brain can describe multiple networks and pub/sub channels, but the reference application currently runs one selected network; general inter-network runtime messaging is incomplete.

These declarations remain useful as language design, validation and future compatibility work, but documentation must not imply that they already execute fully.

Scientific scope

The reference examples are controlled demonstrations, not proof of biological fidelity, general intelligence, automotive safety or superiority to established neural architectures. Strong claims require repeated seeds, uncertainty estimates, external baselines and independently reproducible data.

Compatibility policy before 1.0

Breaking changes are permitted when they remove accidental behavior or make semantics explicit. Models that relied on random bias, implicit plasticity or undocumented defaults may change. Version models, datasets, saved networks and runtime together. Migration notes should accompany every public release.