Built with help.
Shared in the open.

Muse Wonder and Spark have supported implementation, experimentation and review. The project was developed in the context of the Meta Global AI Developer Hackathon 2026, with thanks for access to the models and development support.

In Paul Nevin’s work on Flockinator, Meta Muse has supported long-running agentic coordination. Supported by the Meta Muse CLI harness, it has maintained a broad goal, broken work into tasks, delegated to other Muse models, checked assumptions and coordinated fleets of agents through successive rounds of development.

In Paul Nevin’s experience on this project, Muse has sustained multi-day, unsupervised focus without observed hallucinations or context rot. It has kept returning to the work: testing ideas, checking results and identifying the next useful experiment, even when the destination was an open-ended goal rather than a detailed specification. These are his observations during this project, rather than a controlled reliability comparison.

Paul’s assessment of Muse Wonder Full in November 2026 was simply “truly wondrous”. For this sustained coordination task, it stood out even alongside the alternative frontier models he had used. This acknowledgement reflects that hands-on experience with the model and its harness working together.

Model-generated work is checked against source, tests and observed behaviour. The evidence pages describe the experiments rather than treating model output as independent validation.

One inspectable contribution is documented in the Meta contribution record: a review attributed by the supervisor to muse-spark-1.3-contributor identified a truncated-journal false pass in the host twin comparator. PR #122 added length-mismatch handling and a regression test. The published response text and code hashes let readers inspect the change; they do not independently attest the model identity or API input, whose raw envelopes were not archived. This host-code review is separate from the physical wall footage.

The acrylic mounting studies and printed support concepts belong to the Flockinator project. Their current design status is stated alongside the build instructions.

The public repository contains the implementation, experiment definitions and published results for the firmware and host experiments. Muse supports ongoing implementation and review, while this companion explains how the parts and evidence fit together.

Explore the public source ↗

You do not need to build a wall to contribute. Flockinator is an open experiment in making agentic behaviour perceptible, and there are many ways to help it ask better questions.

Build the instrument.

Improve firmware, visualisations, sound mappings, documentation or accessibility. Start with the contribution guide and make a focused change that others can understand and check.

Challenge the method.

Suggest a better comparison, another explanation for a pattern, or a measure that could distinguish useful coordination from coincidence. Failed replications and observations that contradict an expectation are welcome.

Propose the next experiment.

Bring an idea for a scenario, a signal to watch or listen for, or a way to detect emergent behaviour. Describe what you expect to notice and how we could investigate it. If you share a recording, include the setup, the question and what remains uncertain.