This README was written entirely by Claude
The first public research project from Dandelion Engineering.
This project started from a simple question: could software narrow the gap between a cheap consumer EEG and an expensive research-grade one, without changing the hardware? The seed idea — borrowed from how large language models are "distilled" — was to record the same person wearing both kinds of device at once and train a model to map the cheap signal onto the expensive one. No public dataset like that exists, so the team pivoted to the closest question the available public data can answer: can a model take a research-grade recording stripped down to four consumer-like electrodes and reconstruct the channels that were removed? On the primary dataset it can — a small neural network beats the standard classical methods at every level of the evaluation, on people it was never trained on, by a margin that survives conservative statistical correction. On a second, independent dataset the reconstruction transferred but the downstream task did not, so the broader version of the claim is reported as inconclusive, not confirmed. The honest, narrow result is the result. If you want the science in plain language, read the Accessible Piece; if you want it in full, read the Technical Report.
But the science is not the main reason this repository is public. This is the first time we are showing how a Dandelion Engineering research project actually works — one human and two AI agents, a shared file-system framework, and an idea sharpened into a real, reproducible result over many short sessions. That method is what the rest of this README is about.
A complete, public record of one research project: the question, the plan it was held to, every turn it took, the result it reached, and the artifacts that let anyone check it. Nothing has been cleaned up after the fact to look more decisive than it was. The pivot away from the original idea, the points where the plan was amended, and the parts of the claim that did not hold are all here, on purpose. The transparency is the demonstration.
The project was run by Randy Crespo (human director) and two AI agents, Claude and Codex, inside a structure we call a Collaboration Station. The same structure is described in full in Project Details/Project Details.md; this section is the short version.
This is the part worth your time. The science result above is one project; the method below is what produced it, and is what every future Dandelion project will use.
A Dandelion project is run by one human director and a small team of AI agents working across many short sessions. Each session, an agent reads the project's history, picks up where the last one left off, contributes one real piece of work, and writes a clear handoff for the next session. Nothing is lost between sessions because everything lives in the file system — chats, decisions, code, reports. The strategy is time, not money or headcount: small honest sessions stack up into work that would normally need a much larger team. (The reasoning behind this is in Project Details/Project Details.md § Strategy: Time.)
Every project begins by turning a loose idea into something executable. That document is the Claim Sheet (Claim Sheet.md), drafted by the agents, approved by the director, and then read at the start of every session as the project's contract. It has fixed slots so nothing structurally important is left implicit, including:
- The transferable claim — the one sentence the project could say if it succeeded.
- Constraints — here, no budget, a single consumer laptop, and only free, openly licensed data and tools.
- Methods, materials, and evaluation design — the datasets, the baselines, the metrics, the statistical tests.
- The director's verification path — a commitment, made before any results exist, to build an artifact that lets a non-expert check the result with their own eyes.
- Pre-declared success, failure, and inconclusive shapes — written down before any result is seen, so a partial win cannot later be reported as a full one and a failure that looks like a success gets caught.
Because the Claim Sheet can get technical, every project also produces an Accessible Claim Sheet (Accessible Claim Sheet.md) — the same commitments in plain language, so the director always has a clean grip on what the project promised.
Real research surfaces things the plan did not anticipate. When that happens, the agent who finds it writes an amendment: what was found, why it changes the path, and what the new success/failure shapes are. Amendments are appended to the Claim Sheet, dated, never written over the original — so the plan stays honest to the evidence and the trail of every turn stays visible to anyone reading later.
This project used that protocol for real. The biggest turn is the late decision not to run one of the originally planned sanity-check datasets. You can read the amendments in the Claim Sheet's Amendments section.
A Dandelion project is finished when the work is shippable, which means three artifacts exist:
- A Technical Report (
Technical Report/main.pdf) — rigorous, complete, written in LaTeX for a researcher in the field. - An Accessible Piece (
Accessible Piece/AccessiblePiece.md) — the same project written for someone with no technical background, honest and free of jargon walls. - A Reproducibility Packet (
Reproducibility Packet/) — code, configs, data references, and a README that lets an outside reader reproduce the result on their own machine without contacting anyone on the team.
The director's verification artifact lives inside the Reproducibility Packet, so anyone who downloads it can verify the work the same way the director does — not just read a claim about it.
The director approves the Claim Sheet and then steps back; the agents own the work from there. Two mechanisms keep the director close enough to know the result is real without becoming a bottleneck: a progress report written on a regular cadence in plain language, and the verification artifact named above. Both are written so a generalist can follow what happened and why it matters.
This is a large repository. If you only look at a few things, look at these.
If you want the result, fast. Open the pre-rendered Verification Viewer — a set of static HTML pages, no install required:
Reproducibility Packet/prebuilt_viewer/index.html
Pick any of the 109 held-out participants and ask one question: does the model's reconstruction preserve the eyes-closed alpha bump that the true signal shows? That single check is the heart of the Tier 1 claim, and you can audit it yourself in a browser.
If you want the science in plain language. Accessible Piece/AccessiblePiece.md — what the question was, what was tried, what was found, and what it does and does not mean.
If you want the full technical version. Technical Report/main.pdf — methods, numbers, per-fold breakdowns, the device-grounding analysis, and the statistics.
If you want to reproduce it. Reproducibility Packet/README.md — a copy-paste runbook that reproduces the Tier 1 result end to end in well under an hour on a consumer laptop.
If you came for the collaboration method. This is the most distinctive thing in the repository, and most of it is in the parts a research repo usually hides:
- The contract and its turns:
Claim Sheet.mdand its Amendments section — watch the project commit to a plan, then honestly revise it when the evidence demanded it. - The agent conversations:
chats/— the agents working out the claim sheet, the dataset decisions, and their cross-reviews of each other's work. - The session-by-session trail: each agent's
Session Summaries/andProgress Reports/underagents/— the actual cadence of the work, including the points where something didn't work and was reported as such. - The honest result: the project reached Tier 1 (prototype success on the primary dataset) but reported broader transfer as inconclusive after the same failure pattern repeated on two independent subsets of a second dataset. A team optimizing for appearances would have buried that. We led with it.
A small neural network — about 72,000 parameters, trained on a laptop with no budget — reconstructs high-density EEG from a four-channel consumer-like input on the primary public dataset (PhysioNet MMI) better than the standard classical methods, on participants it never saw, by a margin that holds up under conservative statistical correction. The simulated four-channel input it was trained on is statistically consistent with what a real Muse 2 headband produces on the same kind of resting-state task. That is Tier 1: prototype success.
On a second, independent resting-state dataset (LEMON), probed twice at different sizes, the reconstruction-level result transferred but the downstream eyes-open/eyes-closed classifier did not beat the best classical baseline. The project reports this as broader-transfer inconclusive, not as confirmation or refutation. It is not a claim that a consumer headband can replace a research EEG, and it is not a clinical result. The narrow, honest version is the one that ships.
This project does not redistribute any EEG data. Each dataset is openly published and is fetched by the reader directly from its original host, under its original license. All four carry licenses that permit commercial use. Per-dataset roles, sizes, and download instructions are in Reproducibility Packet/data/README.md; the citations below are the by-the-book forms reconciled into the Technical Report's bibliography.
PhysioNet EEG Motor Movement/Imagery Database — primary training and evaluation dataset (License: ODC-BY 1.0).
Schalk, G. (2009). EEG Motor Movement/Imagery Dataset (version 1.0.0). PhysioNet. https://doi.org/10.13026/C28G6P
The PhysioNet resource itself is also commonly cited via:
Goldberger, A. L., Amaral, L. A. N., Glass, L., Hausdorff, J. M., Ivanov, P. Ch., Mark, R. G., Mietus, J. E., Moody, G. B., Peng, C.-K., & Stanley, H. E. (2000). PhysioBank, PhysioToolkit, and PhysioNet: Components of a new research resource for complex physiologic signals. Circulation, 101(23), e215–e220. https://doi.org/10.1161/01.CIR.101.23.e215
LEMON / Leipzig Mind-Brain-Body dataset — confirmatory resting-state validation (License: CC BY 4.0).
Babayan, A., Erbey, M., Kumral, D., Reinelt, J. D., Reiter, A. M. F., Röbbig, J., Uhlig, M., Anwander, A., Bazin, P.-L., Horstmann, A., Lampe, L., Nikulin, V. V., Okon-Singer, H., Preusser, S., Pampel, A., Rohr, C. S., Sacher, J., Thöne-Otto, A., Trapp, S., Nierhaus, T., et al. (2019). A mind-brain-body dataset of MRI, EEG, cognition, emotion, and peripheral physiology in young and old adults. Scientific Data, 6, 180308. https://doi.org/10.1038/sdata.2018.308
Lee/Ahn 2026 consumer- and research-grade EEG dataset — device-grounding, feature-level only (License: CC BY 4.0).
Lee, Y., Gwon, D., Kim, K., Park, S., Sohn, S., Choi, M., Choi, M., Bae, J.-H., & Ahn, M. (2026). EEG dataset of consumer- and research-grade systems. Scientific Data, 13, 595. https://doi.org/10.1038/s41597-026-06962-5
HEROIC Muse 2 dataset — a Phase-0 planned consumer-only sanity check that was not run before the project closed (the no-run decision is recorded as a Claim Sheet amendment); listed here for completeness (License: CC BY 4.0).
Sugden, R., Campbell, I., Pham-Kim-Nghiem-Phu, V.-L. L., Higazy, R., Dent, E., Edelstein, K., Leon, A., & Diamandis, P. (2024). HEROIC: a platform for remote collection of electroencephalographic data using consumer-grade brain wearables (v2). Zenodo. https://doi.org/10.5281/zenodo.11493897
The Dandelion-authored parts of this project are released under two free, well-known licenses, both of which let others — including companies — reuse the work as long as they credit it:
- Code, scripts, configurations, and runbook material — MIT License, in
LICENSE. - Written and narrative artifacts (this README, the Technical Report, the Accessible Piece, the Claim Sheet and Accessible Claim Sheet, the framework document, the agent workspace records, the Reproducibility Packet's prose, and the verification viewer's user-facing copy) — Creative Commons Attribution 4.0 International (CC BY 4.0), with scope recorded in
LICENSE-DOCS.md.
The raw EEG datasets are not redistributed; each remains under its own original license at its own host (see Datasets and citations above). Dependency licenses are audited in Reproducibility Packet/DEPENDENCY_LICENSES.md. These license files are an engineering record of the project's choices, not legal advice.
To cite this project, use CITATION.cff at the repository root — GitHub renders it as a "Cite this repository" button. When you reuse a result that rests on one of the public datasets, please attribute both this project and the underlying dataset.
Dandelion Engineering is a research and technology company with a single purpose: to do real research, and to turn what we learn from it into affordable technology that materially improves the lives of everyday people. It is not venture-scale and not built to maximize profit — it is a small, deliberate, long-running collaboration between one human director and a team of AI agents, pointed at problems that matter for ordinary people.
This repository is the first public example of that work. The essay that explains the philosophy behind it, and the rest of the project's public links, are listed below.
- The essay: What to do when Everything Changes
- More on the method:
Project Details/Project Details.md
Dandelion Engineering is run by Randy Crespo. If this work or the way it was made resonates with you — whether you're a researcher, an engineer, someone working on adjacent problems, or just curious — I'd genuinely like to hear from you. Thoughtful questions, critique, and ideas for collaboration are all welcome.
- LinkedIn: linkedin.com/in/randy-crespo
- Email: randy@dandelionengineering.com