Choose a task

Ten task families are supported. Each declares inputs, effectors, action cadence, an ordinary horizon and one primary outcome. Choose the opportunity that answers your question before comparing model settings.

Physical control

Task	Question	Native unit
Tracking	Does the observer align with a moving stimulus?	Mean cosine of angular error
Pong	Does the paddle intercept approaching balls?	Hit fraction
CartPole	How long does the pole remain within bounds?	World steps survived

Tracking and Pong run for a fixed horizon. CartPole terminates when its declared bounds are crossed or its step cap is reached.

Capacity probes

Task	Question	Native unit
Delayed cue	Is one cue retained until go?	Response accuracy
Recall interference	Is the first cue retained despite competing cues?	Response accuracy
Delayed XOR	Can two separated cues be combined?	Response accuracy
Evidence accumulation	Is the majority of a pulse stream recovered?	Response accuracy
Context integration	Is the relevant evidence stream selected?	Response accuracy
Temporal order	Is A–B distinguished from B–A?	Response accuracy
Reversal adaptation	Does the response adjust after a hidden mapping changes?	Post-change accuracy

The preset library has 64 named difficulty cells across these seven families. They are configurations, not independent samples or validated abilities. Binary probes finish at the end of their response episode. Reversal is one continuous multi-round session per trial; rounds do not become independent trials.

Inspect a task contract

from brainlesslab import TaskSpec
from brainlesslab.tasks import definition, resolve

task = TaskSpec("delayed_cue", {"delays": (32,)})
print(resolve(task))
print(definition(task))

Expect three inputs, two effectors, one neural frame per world tick, a 48-tick default horizon and outcome key recall_accuracy. Resolution checks options before allocation. The task contract reference lists widths, cadences, horizons, native outcome keys and bounds for all families.

The four-task comparison is an operation across Tracking, Pong, CartPole and delayed cue. It is one protocol, rather than the complete task catalogue. Historical benchmark tables retain their original protocols and source revisions; current availability does not promote them.

Compare conditions within a task using independent blocks and a suitable null. Use performance for execution speed rather than task scores.