Glossary

Each term below has one preferred meaning across the site and the repository. Every entry links to the page that owns the term in full.

Runtime

node — The local neural unit model. A registered NodeSpec owns its builder, parameters, named parameter sets, capabilities, and tags. See Nodes and reservoirs.

reservoir — A runtime population of nodes with wiring and dynamic state. It receives receptor values and returns activity for a body readout. See Nodes and reservoirs.

body — The sensorimotor organisation coupled to a task. A body closes a reservoir through a world, and the reservoir itself stays task-agnostic. See Bodies and interaction.

embodiment — Embodiment is the generic composed body. It assembles the seven component families under stable component IDs. See Interface reference.

agent — One reservoir, one body, and one interaction cycle. See Worlds, tasks, and populations.

ensemble — The shared world lifecycle. The same Ensemble runs one agent and a fixed population. See Worlds, tasks, and populations.

world — The external state an agent acts in, including objects, fields, encounters, and effects. Environment is the abstract type, and TaskWorld is its task-compatibility subtype. See Worlds, tasks, and populations.

task — Registered setup, ports, outcome, anchors, and defaults. A TaskSpec defines how a composition enters an environment and how its declared outcome is measured. See Worlds, tasks, and populations.

composition — The complete runtime composition. A CompositionSpec selects the node, task, body, node count, parameters, task and body options, and interaction timing. See API, plans, and CLI.

interaction cycle — Neural frames executed within one world step. FixedRateCycle(K) samples the world once, runs K neural frames, and emits one command. See Bodies and interaction.

Body components

receptor — One input channel of a reservoir. The body gives receptor channels their sensorimotor meaning. See Bodies and interaction.

effector — One output channel of a reservoir. An actuator turns effector values into a typed command. See Bodies and interaction.

port width — The number of receptor or effector channels a body declares. Port widths must match the reservoir before tick zero. See Bodies and interaction.

geometry — The component family that declares the body’s footprint and geometric bounds. See Interface reference.

sensor — The component family that performs physical sampling of the world. See Interface reference.

encoder — The component family that converts raw sensor values to receptor ports. See Interface reference.

readout — The component family that converts reservoir activity to effector values. See Interface reference.

actuator — The component family that converts effector values to a typed command. See Interface reference.

dynamics — The component family that converts a command and the current motion state to a new motion state. See Interface reference.

physiology — The optional component family for internal feedback, effects, and viability. See Interface reference.

Evaluation

evaluation — The outer trial, seed, reset, and aggregation protocol. An EvaluationSpec is separate from neural frames and world ticks. See Evaluation.

target — One named composition with one evaluation protocol. An operation section names an EvaluationTarget by its id. See Plan file format.

block — The outer repetition unit of an evaluation. A block or trial is usually the independent randomised unit. See Evaluation.

trial — One rollout inside a block. Agents, nodes, samples, and ticks within one world do not multiply the sample size. See Evaluation.

horizon — The number of world ticks in one trial, including warm-up. See Plan file format.

warm-up — Leading ticks that run before recording and scoring. See Evaluation.

scored interval — horizon - warmup, the part of a trial that is recorded and scored. A plan is rejected when this interval falls below the task’s minimum_scored_ticks. See Core catalogue.

construction scope — Whether topology is constructed once per evaluation, once per block, or once per trial. See Evaluation.

reset policy — What is reset between trials. The policies are :full, :body_environment, and :none. Generic operation plans currently support only :full. See Evaluation.

seed stream — A named random stream derived from the evaluation root seed. The default streams are :topology and :world; declare another only when an implementation consumes it. See Evaluation.

aggregation — How declared trial values combine inside one operation. Aggregation does not authorise a cross-task aggregate. See Evaluation.

Operations and records

operation — One research action over targets: a profile, sweep, ablation, evolution, or benchmark. See Operations.

operation plan — The strict TOML file that combines named targets with one operation. See Plan file format.

profile — The operation that runs declared analyses over the raw trials of one target. It produces a descriptive profile. See Operations.

sweep — The operation that evaluates one target across declared parameter cells and pairs seeds across those cells. It produces a development grid. See Operations.

ablation — The operation that compares registered interventions against an implicit baseline. It answers whether a registered mechanism is necessary. See Operations.

benchmark — The operation that reports task-specific statistics and paired contrasts inside each declared case. See Operations.

evolution — The operation that searches registered model coordinates on training targets before held-out evaluation. It is experimental. See API, plans, and CLI.

record — The portable output of one operation. It holds the submitted and resolved plans, authoritative CSV tables, a summary, a seed ledger, provenance, and checksums. See Records and results.

genome — The explicit parameter vector that an operation may search, handled by pack_params and unpack_params. Runtime activations, learned weights, and buffers are not part of it. See Interface reference.

design spec — Evolution.NodeDesignSpec, the reviewed declaration of which model coordinates evolution may change for one node. See API, plans, and CLI.

Scoring and evidence

anchor — A declared reference point for a task outcome. Each anchor carries a value, a kind, and its provenance. A measured anchor also declares its scored interval when that interval affects the measurement. See Core catalogue.

floor — The lower anchor of a task. It is an analytic chance level or a calibrated null control. See Core catalogue.

ceiling — The upper anchor of a task. It is an analytic maximum or a measured reference. See Core catalogue.

normalised score — The task-declared anchor transformation of a raw outcome. A measured anchor can transform only a raw outcome measured over the same scored interval. A normalised score is useful within one task, and it is not a common competence scale. See Core catalogue.

evidence state — The status of an experiment protocol: planned, exploratory, tuned, frozen, confirmed, promoted, or retired. See Experiments and evidence.

Source: docs/WRITING.md, src/core/Specifications.jl, src/core/Interfaces.jl, src/tasks/Tasks.jl, src/operations/Plans.jl.