Platform limits
BrainlessLab tests neural substrates inside declared bodies, tasks, and worlds. The current core starts with Falandays nodes in Tracking, Pong, and Wall. The platform also includes experimental nodes, physical components, collective worlds, analyses, and challenge tasks.
It is not a general robotics simulator, a complete neuroscience model, or an authority that turns a task score into a claim about cognition.
Scientific scope
The platform can support a claim about an implemented mechanism under a recorded protocol. It cannot establish, on its own:
- biological fidelity outside the validated implementation boundary;
- cognition, intelligence, autonomy, or motivation as natural kinds;
- transfer to animals, physical robots, or unmodelled environments;
- optimality outside the searched task and parameter region;
- causal necessity without an intervention and appropriate control;
- equivalence from a non-significant difference;
- criticality from one fitted distribution or scalar estimator;
- general competence from normalised values on different tasks.
Performance can still be informative when it is poor. A task can expose a missing capacity, an incompatible body, an unsuitable timescale, or a parameter trade-off. Check the task opportunity, ports, body, horizon, null, and reset before interpreting failure.
Falandays validation boundary
The :falandays node is numerically fixture-validated against declared trajectories from a
local authors-derived reference implementation. The fixtures cover the tested construction
and update path.
They do not validate every body, task outcome, experimental Falandays variant, analysis, or biological interpretation. Behavioural performance remains a separate empirical result.
Node and learning limits
The canonical Falandays node adapts local targets and recurrent weights during behaviour. It has no reward signal, task-loss backpropagation, or fitted readout. Other registered nodes can have different learning or fixed-weight behaviour.
Therefore, statements about online self-organisation apply to the relevant node, not to every composition that BrainlessLab can run.
Current node models omit most cellular biology. See Nodes and reservoirs for the implemented mechanisms and omissions.
Physical and ecological limits
ObjectWorld currently supports:
- two-dimensional toroidal or walled arenas;
- fixed agent populations;
- static circular objects;
- named analytic scalar fields;
- simplified spectral appearance and illumination;
- typed contact and exposure effects;
- one compatible actuator and dynamics command path per physical body.
It does not provide general meshes, rigid-body collision physics, fluid dynamics, moving non-agent objects, birth, replacement, lineage scheduling, or structural evolution of a component graph.
Body presets named for insects, robots, or aircraft are reusable compositions. They are not validated models of a species or vehicle. A study must justify its sensor geometry, noise, actuator limits, and fields.
Sensing limits
Field probes sample explicit analytic fields. An object bank does not create a field automatically.
Spectral cameras use simplified channels, appearance, illumination, and ray occlusion. They do not provide a wavelength-continuous optical model, diffraction, atmospheric effects, or calibrated hardware response.
SectorVision uses hard angular sectors, finite range, and the nearest target per sector.
Its bearing-shift control preserves sampled values while changing their alignment. This is
one matched control, not a complete set of social nulls.
The shoal-foraging example omits fluid forces, fin mechanics, optic-flow stabilisation, body-body collision response, and species-specific vision.
Evolution limits
EvolutionPlan searches one registered node whose NodeSpec declares a reviewed
Evolution.NodeDesignSpec. Node count and composition structure remain fixed. The included
optimiser is experimental infrastructure. One run does not measure search variance.
DevelopmentSpec varies bounded scalar paths on a fixed embodiment graph. It does not add
or remove components, vary port count, encode runtime state, or schedule developmental
events.
The platform does not yet provide one typed operation that evolves arbitrary node structure, connectivity graph, and embodiment structure together.
Analysis limits
Finite runs and populations constrain every estimator:
- agent-scale statistics need enough independent agents and worlds;
- avalanche fits need enough events and sensitivity checks;
- spectral radius is a linear matrix proxy;
- transfer entropy needs adequate samples, stationarity checks, surrogates, and multiple-pair correction;
- shared input can imitate collective coupling;
- failed, dead, missing, and saturated runs need declared handling.
The Analysis reference gives method-specific checks. Every new task and sampling regime still needs validation.
Workflow limits
BrainlessLab records plans, named seeds, resolved settings, tables, reports, provenance, and checksums. This supports a traceable rerun in a declared environment.
The platform does not yet provide a graphical study builder, automatic preregistration, sealed-data service, permissions sandbox, or automatic evidence promotion. Bit-for-bit claims apply only where explicit fixture tests establish them.
The public run-contribution pipeline accepts public records only. Aim for at most 1 MiB of text-only files in the complete contribution. The hard limit is 5 MiB. Files above this limit and large datasets remain out of scope for this Git-native path.
Direct Ensemble simulation supports heterogeneous bodies and reservoirs. Some
configuration and optimisation paths still require a fixed receptor and effector layout
for the population they construct.
Use Experiments and evidence for the evidence ladder and Extend from another project for the supported extension path.