Who you are, and what to use the Crown for

People arrive at a Crown from four directions, and the traps are different in each case. This is what I would tell each of them, with the pages that go into the detail.

If you have just unboxed a Crown

What to watch out for

• Both scores are probabilities from 0.0 to 1.0, and Neurosity’s documentation says over 0.3 is significant for either one. A 0.35 is a meaningful reading, not a near miss. This is the most common first-week misreading, and I made it too.
• Both scores take up to 16 seconds to initialise, so nothing at the very start of a session means anything at all.
• Your own first minute is settling in. Posterior alpha in my recordings opened around 3 to 4% and then climbed steeply, so an average taken over a whole session is dragged down by the part where you were still getting comfortable.
• A high focus score can be a clenched jaw. The focus band is 30 to 44 Hz, which is where jaw and brow muscle activity lives, and above roughly 20 Hz most of what a scalp electrode records on an awake head is muscle rather than cortex.
• Comparing your numbers against someone else’s is meaningless. No cross-session or cross-person normalisation layer exists on this platform.

Where to go. Never touched an EEG headset? Start here for the four things and the picture. Trust the calm score first for why calm is the score to build a habit around. Throw away the first minute of every recording for reading a session curve. And the five-word scale on what I built, because that vocabulary is what keeps a reading self-relative instead of turning into a grade.

Something you may not have considered. Run the eyes closed test before you trust anything else: 30 seconds with your eyes open, 30 seconds closed, and check that alpha rises. Alpha power falls by roughly 40% going from eyes closed to eyes open in healthy adults (Barry et al. 2007), so it is the cheapest possible check that the fit and the pipeline are both working. If it comes out flat, run it once more before you blame the headset: standard clinical EEG descriptions note that a minority of perfectly healthy adults have a low-voltage background with little measurable alpha (Nayak and Anilkumar, EEG Normal Waveforms, StatPearls).

If you are a developer or an engineer

What to watch out for

• The CSV export has no header row, so line one is already sample zero and a spreadsheet will quietly consume it as column names.
• Slice by sample index, never by time. The timestamps step backwards once or twice per file, with occasional gaps.
• Delta dominates, at 72 to 95% of total power below 4 Hz, so bands only compare as shares of the total and never as absolutes.
• A loud channel is almost always bad contact rather than a busy region. In one of my sessions F5 ran 26 times the median. Judge a channel by how much it moves, not by its DC offset.
• Console-trained models are fitted to one person’s head. The transferable output is the method, not the weights.
• The data leaves the headset two ways, with different failure modes. Local broadcast (OSC, LSL, BrainFlow, no login, dies the moment you leave the Wi-Fi, blocked by firewalls) or cloud (SDK, MCP, WebSockets, login required, works anywhere).

Where to go. Don’t open the raw export in Excel for the export traps. A loud channel isn’t a busy brain for signal quality. Three things this headset can already do for the unused surface area. And what I built for the computed-then-written architecture, if you want a pattern for keeping an interpretation layer auditable.

Something you may not have considered. If you are writing your own spectral analysis, the method behind every number on this site was a 1 to 45 Hz zero-phase Butterworth filter followed by Welch spectra in four-second windows. Copy it or argue with it, but start from something stated. And before you trust anything above 20 Hz on a dry-electrode headset, read the paralysis study: scalp recordings taken before and during neuromuscular blockade showed that above roughly 20 Hz, most of what survives on a normal awake head is muscle (Whitham et al. 2007).

If you are a researcher or a meditation practitioner

What to watch out for

• The alpha and theta reading is the defensible part of this device, and it is defensible because the literature is behind it. Meditation-related increases in posterior alpha and in theta are well established (Lee et al. 2018; Lagopoulos et al. 2009; Lomas et al. 2015). Everything else this hardware offers is on weaker ground than that.
• Eight dry electrodes, with no midline, no temporal and no true prefrontal coverage. That rules out foot imagery outright and constrains anything that needs frontal midline theta.
• Nothing here is diagnostic or clinical. No reading should ever be described to a client as good or bad, only as usual or unusual for that person on that day.
• Read theta next to alpha, never alpha alone. Theta rises with mental fatigue during imagery tasks (Li et al. 2024), and the transition from wake into the first stage of sleep is described as a change in frequency from alpha to theta (StatPearls, EEG Normal Sleep). A rising calm score is not on its own evidence of a good session.

Where to go. Trust the calm score first for why calm is the more trustworthy of the two scores. Throw away the first minute of every recording for reading a session curve honestly. Build on alpha first for the evidence under the calm score and the one-minute test that checks your setup can see it, and what I built for why the output is sentences rather than numbers, including the scope limits on what affect labelling does and does not show.

Something you may not have considered. One ten-minute recording would turn the central claim on this site from an inference into a demonstrated result: a single session alternating deliberate jaw clenching with genuine concentration. It is the next thing I would run, and I have not run it yet. If focus spikes on the clenches, the argument for trusting calm over focus stops being my opinion. I would rather someone ran it than took my word for it.

If you are evaluating this as a product

What to watch out for

• Platform dependency is the structural fact to plan around. Anything built here is an accessory to Neurosity’s hardware and Neurosity’s console, and none of that layer is under your control.
• Console-trained trials are not inspectable. No trial list, no per-trial scores, no export, and nothing the MCP server returns. That is a hard ceiling on any product that wants to show a user their own training.
• Every multi-label experiment in my log recorded zero trials. All 219 sit in single-label experiments. Any roadmap that assumes multi-class training works in the console today needs to check that assumption before it goes on a slide.
• A live score on screen is a design trap. The closest verified evidence is about instructed relaxation under load, not about EEG feedback, so I label the step from that evidence to a screen layout as my own design reasoning rather than a finding. I demoted the live view on that basis.
• Inferring emotions is a regulated act in the EU, not a feature category. Article 5(1)(f) of the AI Act puts emotion inference in workplaces and education institutions on the list of prohibited practices, with a narrow exception for medical or safety reasons.

Where the regulation is going. A calm score is an inference about a person’s internal state, which puts it inside a category the EU has already fenced off. Article 5 of the AI Act prohibits placing on the market, putting into service, or using AI systems to infer emotions of a natural person in the areas of workplace and education institutions, except for medical or safety reasons. That is not a compliance checklist item. It sits in the Act’s short list of outright prohibited practices. The moment a Crown product moves from “I read my own sessions” toward “we read our team’s sessions,” it is walking toward that line and toward every consent problem underneath it, and brain data sits about as close to the person as data gets. This turns an architectural choice into a strategic one: local, single-user, no server, no account, nothing leaving the machine costs very little to build and removes an entire category of risk. Local-only and single-user is a selling point, not a limitation to apologise for.

Where to go. Once you train the headset, those trials are gone for the capability ceiling. Three things this headset can already do for the surface area nobody has used. And the case study for why the pivot is the finding.

Something you may not have considered. Two documented platform capabilities have no consumer product built on them: two haptic motors at P7 and P8 accepting up to seven named effects per command, and OSC Perform Mode, which streams interpretations four times a second over a 16-second sliding window. Haptics in particular is the obvious way to close a feedback loop without putting a number on a screen, which sidesteps the live-score problem entirely rather than managing it.

Sources

My own data: three recorded console sessions (248,320 samples), an audit of seventeen training experiments, the crown-debrief repository, and the Notion console and sensor references.

• Neurosity documentation: Focus and Calm
• Barry, R. J. et al. (2007). EEG differences between eyes-closed and eyes-open resting conditions. Clinical Neurophysiology 118(12):2765-2773.
• Whitham, E. M. et al. (2007). Scalp electrical recording during paralysis. Clinical Neurophysiology 118(8):1877-1888.
• Lee, D. J. et al. (2018). Review of the Neural Oscillations Underlying Meditation. Frontiers in Neuroscience 12:178.
• Lagopoulos, J. et al. (2009). Increased Theta and Alpha EEG Activity During Nondirective Meditation. Journal of Alternative and Complementary Medicine 15(11):1187-1192.
• Lomas, T., Ivtzan, I., and Fu, C. H. Y. (2015). A systematic review of the neurophysiology of mindfulness on EEG oscillations. Neuroscience and Biobehavioral Reviews 57:401-410.
• Li, T. et al. (2024). Research on mental fatigue during long-term motor imagery. Scientific Reports 14:18454.
• Nayak, C. S., Sankari, A., and Anilkumar, A. C. EEG Normal Sleep. StatPearls.
• Regulation (EU) 2024/1689 (EU AI Act), Article 5(1)(f).