Delayed cue

What is remembered?

The task presents either left or right cue for eight ticks, waits a chosen delay, then presents go for an eight-tick response. Its three receptors are left cue, right cue and go; its two effectors choose left or right. There is one neural frame per world tick.

cue (8) → delay (chosen from 8, 32, 128) → go / response (8)

Labels are balanced binary draws from the world stream. Effector values accumulate only during response; ties choose left. The default horizon 144 contains the longest default delay, while a completed short-delay episode ends sooner.

from brainlesslab import CompositionSpec, NodeSpec, TaskSpec, profile, resolve

target = resolve(profile(CompositionSpec(
    node=NodeSpec("sorn"),
    task=TaskSpec("delayed_cue", options={"delays": [32]}),
    count=32,
))).targets[0]
print(target.horizon, target.n_inputs, target.outcome_key)

Expect 48 3 recall_accuracy. The horizon follows cue plus delay plus response. Warm-up is zero and the complete episode must fit inside the declared horizon.

Choose the visibility control

cue_mode="transient" removes the cue after its initial window. "persistent" leaves it visible through response, allowing a memoryless solution. "hidden" removes it throughout. These choices preserve the seeded episode choices but answer different questions.

recall_accuracy is zero or one for a trial and an accuracy across trials. Analytic chance is 0.5; the signed adjustment is 2 * raw - 1. The observation-based oracle remembers the initial cue. Memoryless, last-cue, constant and random-action references test narrower shortcuts.

Success at one short delay does not establish a retention curve. Use recall interference for declared difficulty cells, then offline decoding to ask whether emitted activity contains recoverable cue information.