BrainlessLab

Put a neural design in a task

Choose a question with an observable outcome: alignment, interception, balance or a delayed response. Each task retains its own native units.

Task illustration · Scripted diagram, not a recorded simulation.

Tracking

View docs

Does the observer align with a moving stimulus? Outcome: mean cosine of angular error.

The task guide covers three physical tasks and seven capacity-probe families, with 64 preset probe cells. Explore the Falandays and SORN models to see their equations and local adaptation rules.

Declare, compare and inspect

1. Declare a design. Specify the model, task, node count and randomised repetitions in a plan.
2. Compare conditions. Keep the task opportunity and world streams matched while changing a declared mechanism.
3. Inspect the record. Read resolved settings, stream identities, trial status and task outcomes before interpreting a summary.

Start by checking the included delayed-cue plan:

uv run brainlesslab check python/plans/delayed-cue.toml

From a repository checkout with the package installed, this prints the resolved configuration without running a simulation. Run your first simulation for installation, execution and a checksum check of the saved record.

Then create a reproducible profile to measure variation across independent blocks, or compare conditions to test a declared change. The system map explains how Python constructs the system and Quadrants advances neural and world state in bounded batches.

What’s it like to be a neuron?

Start with the local view: inputs, activity and adjustment. Then ask what changes when many simple units interact.

BrainlessLab gives this question an experimental form. Specify a neural reservoir, place it in a declared task, and compare what happens. The opening visual introduces the cellular perspective; it is not a BrainlessLab simulation or evidence about a neuron’s experience.

Research

Our first two studies ask when local adaptation helps task performance. The initialisation study maps the effects of starting weights and plasticity across four tasks. The Falandays–SORN follow-up examines how those effects depend on recurrent strength and the adaptation mechanism.

Read the articles or contribute an investigation.

Keep software and evidence separate

A completed run establishes execution. A task score does not establish cognition, biological fidelity or general capability. Availability describes software, not scientific confirmation.

Earlier Falandays validation covers declared reference trajectories at its original implementation boundary; see version history. Python equation conformance, backend qualification and observed task performance need separate checks. Accepted records retain their source revisions and evidence status.

Read the platform limits and the research workflow for the current scope and how to interpret results.

People and support

Created by Ben Gaskin, Polyphony Bruna, Ian Jackson and William O’Hearn.

A project of the Diverse Intelligences Summer Institute 2026, Geneva, New York. DISI is supported by the John Templeton Foundation.