Capacity probes

Which question does each probe ask?

The capacity library contains 64 development cells across delayed cue and six additional task families. They share binary responses but differ in the observations that make the answer recoverable.

Family	Question	Preset cells
Delayed cue	Retain one cue through a delay	5
Recall interference	Retain the first cue despite competing cues	6
Delayed XOR	Combine two separated cues	4
Evidence accumulation	Recover the majority of a pulse stream	9
Context integration	Select the relevant evidence stream	36
Temporal order	Distinguish A–B from B–A	3
Reversal adaptation	Adjust after a hidden mapping changes	1

from collections import Counter
from brainlesslab.tasks import capacity_probe_presets

cells = capacity_probe_presets()
print(len(cells))
print(dict(Counter(cell.task for cell in cells)))

Expect 64 and the counts above. Each preset stores an ID, task kind, complete task options and horizon. It is a difficulty cell, not an independent sample.

Start with one declared cell

Construct a TaskSpec from the preset, choose a model and resolve a profile. Use fixed wiring within each independent block and full trial resets. Then compare native response accuracy with references that expose shortcuts: persistent input, last observation, constant choice or observation-based oracle.

Binary chance is 0.5. Scores below chance retain negative null-adjusted values. The reversal outcome uses its post-change window; it is not the average accuracy of every round.

These are available experimental tasks, not a confirmed capacity assessment. The historical planned study retains its own source and evidence status. Current software does not execute or promote that study.

Continue with a family page, paired comparisons and offline decoding.