Glossary

Use these meanings in plans, code reviews and reports. Historical pages may use different implementation terms; their source revisions own those definitions.

Composition and execution

node — One neural unit within a selected model. NodeSpec chooses the model kind and its parameters. Models describe each mechanism.

reservoir — The constructed neural population, wiring and dynamic state that receives inputs and emits activity. Nodes and reservoirs owns the contract.

composition — A CompositionSpec selecting node model, task, count and input gain. Execution placement is separate.

task — A declared observation, action, world and scoring contract selected by TaskSpec. See tasks.

world — The external task state for one trial. Different worlds in a batch are independent slots, not a shared ecology.

body — The task’s sensorimotor interface. The current package implements fixed task interfaces, not a general body-component registry. See bodies and interaction.

agent — One controller interacting with one task world. Agents or ticks inside a world do not provide extra independent replications.

receptor — One encoded input channel supplied to the model. Its meaning and scale belong to the task contract.

effector — One model readout channel supplied to the task’s action rule. It is distinct from a scored outcome.

port width — The number of input or effector channels. Resolution checks widths before allocation.

readout — The selected-node activity mean used to form an effector. An empty selection returns zero.

input gain — The nonnegative composition scale applied after task encoding and before the model step. It is separate from task sensory gain and model input-weight parameters.

interaction cycle — Neural frames within one world tick. CartPole uses 24; the other current tasks use one.

numerical policy — Arithmetic dtype, fast-math rule, reduction order and RNG policy, declared by NumericalPolicy.

execution specification — Backend, batching, resource budgets, threads, cache and recording, declared by ExecutionSpec. See performance.

Repetition and randomness

evaluation — The repetition, construction, seed and scoring protocol declared by EvaluationSpec.

target — A named EvaluationTarget containing one composition, controller, stream identities and any interventions. The plan supplies its evaluation protocol.

block — An independent outer repetition, normally with one constructed wiring and several reset trials.

trial — One rollout within a block. Reversal uses a whole multi-round session as one trial.

horizon — Maximum world ticks in a trial, including warm-up. Early ordinary termination may produce fewer completed ticks.

warm-up — Leading ticks that advance state before scoring. It is not a neural reset and is allowed only by the relevant task protocol.

scoring window — The declared part of a trial used for the primary outcome. Reversal instead uses 16 post-change rounds.

construction scope — Whether construction is shared per evaluation, per block or per trial. Offline decoding requires block scope.

full reset — Restoration of initial world and neural state, including weights, masks, thresholds or targets and learning flags, between trials. This is the supported trial reset.

seed stream — A purpose-specific random stream derived from declared identities and the root seed.

seed partition — The declared bank for development, calibration, training, tuning, pilot, confirmation or robustness. Separate purposes require disjoint seeds.

aggregation — Combining trials within a block, then independent block means. It does not authorise a cross-task average. See evaluation.

Plans and records

operation — Profile, sweep, ablation, benchmark or calibration, represented by the shared Plan schema.

plan — The scientific question, version, named targets, operation, protocol, numerics and evidence state. See plan format.

profile — A one-target descriptive evaluation plan.

sweep — A plan of explicitly declared parameter cells.

ablation — A baseline comparison with declared interventions. The resulting contrast is specific to that intervention and protocol.

benchmark — Task performance across declared cases and references. Runtime speed is a separate measurement.

record — A versioned portable bundle of request, resolved plan, trials, seed ledger, summary, provenance and checksums. See records and results.

Scores and evidence

reference — A declared analytic value or measured control used to interpret a task outcome. Its policy and provenance matter.

null mean — The analytic chance or frozen empirical reference mean. It is not a lower bound on observed performance.

upper bound — The mathematical maximum used by a task’s adjustment. It is not proof that an implemented oracle attains it.

null-adjusted score — (raw - null_mean) / (upper_bound - null_mean). It may be negative and is not a common competence scale.

calibration — A frozen task-specific null distribution or declared analytic chance, matched by protocol signature. See analysis.

decoding — Offline label recovery from recorded activity using whole-trial fit, validation and evaluation splits. It does not supply labels to the online controller.

evidence state — Planned, exploratory, tuned, frozen, confirmed, promoted or retired protocol status. Execution completion does not advance it automatically. See experiments and evidence.

Historical vocabulary

Embodiment, Ensemble, geometry, sensor, encoder, actuator, dynamics, physiology, model profile, evolution, genome and design spec occur in the archived implementation’s generic composition and search interfaces. They do not name additional public Python APIs. Their original meanings remain at the historical source boundary.