Can a $70 board tell me I'm wrong?

Generated from this site's git history. with Claude means the commit carries a co-author line for an Anthropic model: an LLM took part in that change, not necessarily in every word of it, and git cannot say which. The same history on GitHub →

1. v9 2026-09-25 13:41 $70 board post: link the public olimex-shield repo, state the scale's provenance Joshua Anderton with Claude Opus 5.5 +91 −2 words

   - **Not explained by electrode prep.** The one capture the app scored good-contact (alpha SNR 36) has the highest value, 750 nV. The one it flagged no-scalp-contact gives 238 nV. The spread is unexplained.

   - **Not absolutely calibrated.** Counts convert at 7.9 µV/count, measuredderived from the shield's built-in calibration signal at its nominal amplitude; no external signal has been injected to check it. The datasheet implies 1.72. That 4.6× disagreement is unresolved, and every nV figure here inherits it.

   ## Broadband RMS overstates in-band noise 3.5–24×

   Board and microcontroller about C$75; the whole rig about C$106, at listing prices on 2026-09-13.

   The firmware, the bridge and the recording workbench are public: [olimex-shield](https://github.com/polarizetech/olimex-shield).

   | parameter | value |

   | gain | ≈2848, set by a hardware trimmer |

   | scale | 7.9 µV/count measuredfrom the built-in cal signal (nominal amplitude, not externally checked); datasheet implies 1.72 |

   | input-referred noise | **not measured** |

   A paragraph noting that beta-band and dichotic runs had no producible records was removed rather than kept as anecdote. The post was restructured to lead with the instrument characterisation.

   ## Update — 2026-09-25

   The rig's code is now public and linked from the parts list. The scale is described more exactly:

   7.9 µV/count comes from the shield's built-in calibration signal at its nominal amplitude, not from an externally injected signal.

   Commit e43fbba on GitHub →

2. v8 2026-09-25 08:23 Use polarize-ui for the design system and the blog components Joshua Anderton with Claude Opus 5.5 +1 −0 words

   ## The rig

   <div class="ui-parts"> {% include part.html name="Olimex SHIELD-EKG-EMG" role="Analog front end. 0.16–40 Hz, ahead of the converter." price="C$34.02" href="https://www.digikey.ca/en/products/detail/olimex-ltd/SHIELD-EKG-EMG/3471383" img="/assets/posts/rig/olimex-shield-ekg-emg.jpg" credit="DigiKey" %} {% include part.html name="Olimex SHIELD-EKG-EMG-PRO" role="Passive electrode cable for the shield: 3.5 mm jack to three snap leads." price="C$16.91" href="https://www.digikey.ca/en/products/detail/olimex-ltd/SHIELD-EKG-EMG-PRO/5247133" img="/assets/posts/rig/olimex-shield-ekg-emg-pro-cable.jpg" credit="DigiKey" %} {% include part.html name="Arduino Uno R3" role="10-bit converter and USB serial." price="C$40.99" href="https://www.amazon.ca/dp/B008GRTSV6" img="/assets/posts/rig/arduino-uno-r3.jpg" credit="Amazon.ca" %} {% include part.html name="AVCOO snap TENS pads, 20" role="The electrodes. Self-adhesive, reusable gel." price="C$13.99" href="https://www.amazon.ca/dp/B09R1WSVJS" img="/assets/posts/rig/avcoo-snap-tens-pads.jpg" credit="Amazon.ca" %} </div>

   Board and microcontroller about C$75; the whole rig about C$106, at listing prices on 2026-09-13.

   Commit ded42f5 on GitHub →

3. v7 2026-09-13 17:08 Post: the $70 board, rebuilt around what the rig actually measures Joshua Anderton with Claude Opus 5 +1685 −1264 words

   description: >-

   An n-of-1 EEG rig, an analog ceiling that removes the signal I was designing

   for, and a question I can't answer alone.

   A single-channel EEG rig characterised in numbers: in-band 40 Hz power across

   six captures, why broadband RMS is the wrong figure to specify a rig by, and a

   proposed test of analog bandwidth as a validation axis.

   date: 2026-08-25 09:00:00 -0600

   citations:

   - krigolson2017

   - boere2023

   - goldwyn2014

   - iaccarino2016

   - martorell2019

   - goldwyn2014

   image: /assets/posts/eeg-recording-cap.jpg

   ---

   Solo research programme, about two months old. The honest next step is to have someone tell me the instrument is inadequate.

   Written for people who validate portable EEG hardware. The ask is at the bottom and it is small.

   n-of-1 instrument characterisation. Written for people who validate low-cost EEG. One ask, after the rig sections. The auditory question that prompted the work is last.

   ## In-band power at 40 Hz: median 157 nV, 11.7× spread

   Six captures. One rig, one montage, one person. Amplitude at 40 Hz over 60 s of pre-stimulus baseline, estimated from neighbouring frequency bins.

   | | |

   |---|---|

   | median | **157 nV** |

   | range | **64–750 nV** (11.7×) |

   | captures with ≥20 s usable baseline | 6 of 12 |

   | baseline segments whose noise integrates with dwell | 5 of 9; the other 4 are drift-limited |

   What this figure is not:

   - **Not instrument noise.** No shorted-input capture exists. This is in-band power of a scalp recording, EEG and scalp EMG included.

   - **Not explained by electrode prep.** The one capture the app scored good-contact (alpha SNR 36) has the highest value, 750 nV. The one it flagged no-scalp-contact gives 238 nV. The spread is unexplained.

   - **Not absolutely calibrated.** Counts convert at 7.9 µV/count, measured from the shield's calibration signal. The datasheet implies 1.72. That 4.6× disagreement is unresolved, and every nV figure here inherits it.

   ## Broadband RMS overstates in-band noise 3.5–24×

   Specifying a rig by broadband RMS assumes white noise. EEG is drift-dominated. On the same six captures, a white-noise estimate built from broadband RMS put the 40 Hz noise **3.5–24× too high in amplitude** (median 15×), which is **12–594× too long in required dwell**.

   The error runs one direction: it declares runnable measurements unrunnable. A low-cost rig costed from a broadband figure gets ruled out of experiments it could run. Measure noise at the frequency of interest.

   Scope: six captures, one person, one rig. The direction transfers. The ratio may not.

   ## The rig

   Single-channel Olimex SHIELD-EKG/EMG on an Arduino. **Active Oz, reference Cz, ground mastoid** — a posterior montage chosen for alpha. 250 Hz sampling, 10-bit ADC at roughly

   7.9 µV per count, **~40 Hz analog low-pass**. Closed-back headphones. Python doing Welch PSD with FOOOF for aperiodic separation; Web Audio generating the stimulus, so the delivered envelope is measured rather than assumed.

   Microvolt EEG on a 10-bit converter is a few counts. Quantisation-limited and low-SNR by construction.

   ## The alternatives, and why this board

   <div class="parts"> {% include part.html name="Olimex SHIELD-EKG-EMG" role="Analog front end. 0.16–40 Hz, ahead of the converter." price="C$34.02" href="https://www.digikey.ca/en/products/detail/olimex-ltd/SHIELD-EKG-EMG/3471383" img="/assets/posts/rig/olimex-shield-ekg-emg.jpg" credit="DigiKey" %} {% include part.html name="Olimex SHIELD-EKG-EMG-PRO" role="Passive electrode cable for the shield: 3.5 mm jack to three snap leads." price="C$16.91" href="https://www.digikey.ca/en/products/detail/olimex-ltd/SHIELD-EKG-EMG-PRO/5247133" img="/assets/posts/rig/olimex-shield-ekg-emg-pro-cable.jpg" credit="DigiKey" %} {% include part.html name="Arduino Uno R3" role="10-bit converter and USB serial." price="C$40.99" href="https://www.amazon.ca/dp/B008GRTSV6" img="/assets/posts/rig/arduino-uno-r3.jpg" credit="Amazon.ca" %} {% include part.html name="AVCOO snap TENS pads, 20" role="The electrodes. Self-adhesive, reusable gel." price="C$13.99" href="https://www.amazon.ca/dp/B09R1WSVJS" img="/assets/posts/rig/avcoo-snap-tens-pads.jpg" credit="Amazon.ca" %} </div>

   Board and microcontroller about C$75; the whole rig about C$106, at listing prices on 2026-09-13.

   | parameter | value |

   |---|---|

   | channels | 1 differential |

   | montage | active Oz, reference Cz, ground mastoid; chosen for posterior alpha |

   | electrodes | self-adhesive snap **TENS pads** — made for stimulation on skin, not recording through hair |

   | impedance | **not measured** |

   | analog band | 0.16–40 Hz, in the shield, ahead of the converter |

   | converter | 10-bit |

   | sample rate | 250 Hz in three captures, 256 Hz in three |

   | gain | ≈2848, set by a hardware trimmer |

   | scale | 7.9 µV/count measured; datasheet implies 1.72 |

   | input-referred noise | **not measured** |

   | stimulus timing | **not measured** (below) |

   | analysis | Python; Welch PSD; FOOOF for aperiodic separation |

   BoughtChosen on hearsay — forum consensus that thea expensivecostlier board buys a quieter front end and convenience, not a capability you cannot otherwise reach. **That is not a measurement.** No matched-conditions noise-floor comparison of the two has been seen here, and none was run. Landscape, becausechecked running one means owning both.2026-08-25:

   Landscape, checked 2026-08-25:

   | | channels | price |

   |---|---|---|

   | Olimex SHIELD-EKG/EMG *(whatthis I haverig)* | 1 | **€19.95** |

   | Olimex EEG-SMT | 2 | €99.00 |

   | OpenBCI Galea | — | $60,504 |

   The gap between rows one and three is the question. Two failure modes that cannot be ruled out from here: the noise-floor claim may hold for alpha and fail for anything smaller; or the wrong axis was optimised entirely, in which case **neither** Olimex board solves the problem and cheap-versus-expensive was never the interesting comparison.

   ## TheTiming: noiseno floornumber exists

   Median in-band noise at 40 Hz over 60 s: **238 nV**, ranging 65 to 750 nV across eight sessions. An **11× spread on the same hardware in the same room** — electrode prep dominates everything the software does.

   One correction worth passing on. A broadband RMS figure had been carried as if it were the noise floor at the frequency of interest. Real EEG is drift-dominated, so broadband RMS

   **overstated 40 Hz noise by 20–90× in amplitude.** That error declares a runnable measurement unrunnable, which quietly kills experiments that should have been attempted.

   Stimulus from Web Audio in a browser. EEG over USB serial. No loopback channel, no shared clock.

   **No latency or jitter figure exists for this rig.**

   For scale: a consumer Bluetooth headband, tested with 5,000 pulses looped through its auxiliary input, measured a mean lag of 40 ms (±20 ms), with fewer than ten extreme latencies inflating the spread.

   {% include source.html key="krigolson2017" %}

   The serial stream carries a second column. It has not been confirmed as an independent input and may be a floating pin. Until it is, there is no known-good channel to loop the audio into.

   ## Retention

   | stage | kept |

   |---|---|

   | captures with ≥20 s usable pre-stimulus baseline | 6 of 12 |

   | of those, protocol run to completion | 3 of 6 — two stopped in baseline, one in drive |

   | resting-alpha resonance profiles scored usable by the app | 1 of 6 |

   The five unusable profiles: drift-dominated baseline (1), no alpha peak above the 1/f floor (3), inconclusive at 12 Hz (1). The usable one peaks at 10 Hz with 2.8 dB SNR.

   The headband validation above reports loss the same way: participants lost fell to one in twenty once its assistants had experience fitting the headband.

   ## WhatAlpha haswas beenthe runwrong calibration

   The chain recovers resting alpha: SNR 36 in the good-contact capture, and one usable resonance profile of six. Alpha is the weakest available check: present at rest, needs no stimulus timing, exercises nothing between stimulus and record.

   The planned replacement is a visual oddball. The headband validation used one — 25% oddballs, three blocks of 40 trials, filtered 0.1–15 Hz — and recovered the N200 and P300 on consumer hardware. Its task code and a data converter are published. A two-channel forehead-mounted system has been examined on the same two components.

   {% include source.html key="boere2023" %}

   An oddball tests stimulus, timing and record together, against a published yardstick. It needs a montage change first: the lab-grade arm of that validation measured the P300 at Pz. Oz–Cz is not a P300 montage.

   ## The ask: analog bandwidth as a validation axis

   Low-cost EEG validation has a channel axis: how few electrodes still recover ERPs. The two-channel work above sits on it. The validation read for this post targets the N200, P300 and reward positivity, and filtered its consumer arm at 0.1–15 Hz. For those components a 40 Hz analog ceiling costs nothing.

   The other axis is analog bandwidth: how narrow the front end can be before a given evoked response is lost. None of the sources read for this post examine it. No systematic search was done, so it is not claimed to be untested. This rig sits on its edge: 40 Hz is the filter corner, and anything faster is gone before digitisation.

   **The proposal.** A bandwidth sweep on existing lab-grade recordings. An oddball and a 40 Hz auditory steady-state response, each re-filtered with a causal low-pass at a series of cutoffs — for example 15, 30, 40, 60 and 100 Hz — reporting at each cutoff: component amplitude and latency, detection rate, and participants retained. A causal filter approximates an analog front end's phase lag; a zero-phase filter would not. No new participants. The analysis would be done here, on the lab's data, under the lab's protocol.

   **Deliverable:** a cutoff-versus-recovery table for components spanning the band.

   ## The auditory question this started from

   IsochronicThe alpha runs were isochronic tones at the alpha resonant frequency, looking for a measurablean aftereffect. The question underneath is auditory.

   Reported as the data reads, not as remembered. **Of six resonance profiles on file, five are `inconclusive`** — flagged by the app as drift-dominated, or with no alpha peak above the 1/f floor, or with a peak below the alpha band entirely. The one marked `strong` sits at 10 Hz with 2.8 dB SNR. The best recorded run gives a resonant frequency of 11 Hz against a peak of 9.77, a divergence of 1.23 Hz, and its own verdict is **`LOW QUALITY / no valid comparison`**.

   So: the chain demonstrably recovers alpha, and that counts as **rig validation and nothing else** — evidence that board, electrode, amplifier and analysis code together find a real, expected signal. A calibration, not a finding. n = 1, unblinded, no sham for arousal or attention, effect measured *during* drive rather than after, and acute entrainment of an ongoing oscillation is among the least surprising results in the auditory literature.

   A beta-band version and dichotic drive experiments were also run. **No numbers are reported for either, because the records cannot currently be produced.** The beta app was retired and its results are not where expected. Until the data can be put in front of you, those are anecdotes.

   ## The published precedent

   The medial superior olive is the first place in the ascending pathway where the two ears are compared. Phase-locked activity there produces a far-field potential, the neurophonic.

   {% include source.html key="goldwyn2014" %}

   Engagement was with the released model, not the paper's full text. The published code has **no sodium conductance and no axon** — an h-current and a low-threshold potassium current only — and reproduces the far-field without spike-generating machinery. A matched far-field licenses *drive arrived at the nucleus*. It does not license *cells crossed threshold*.

   The defensible question: can a designed acoustic protocol put a specified structure into a recorded evoked response that tracks a parameter sweep and is distinguishable from a matched-energy tone? A delivery-verification methods result.

   This rig cannot run it. The target sits at carrier rates in the hundreds of hertz; the analog filter removes it before the converter. A 100 Hz fundamental has **zero harmonics in the passband**, and no averaging recovers what was never digitised. Three consecutive workstreams were designed against that ceiling before it was checked.

   Under dichotic drive the response envelope can modulate at the difference frequency with no acoustic envelope at either ear. That appears as **sidebands around the carrier**, not energy at the difference frequency. Whether a single channel can recover that structure is a hardware question.

   The nearest thing in the literature to thepublished shape being chased is the 40 Hz sensory drive work in mouse modelsmice, later extended to combined auditory and visual stimulation.

   {% include source.html key="iaccarino2016" %} {% include source.html key="martorell2019" %}

   Cited as **precedent that the shape exists**, not as support. Neither was read at full text. Four differences bound how far they travel: mouse versus human, invasive tissue readout versus scalp EEG, 40 Hz gamma versus alpha, cortex and hippocampus versus a brainstem nucleus. **The readout difference is binding.** Those studies read the consequence in tissue. Nothing available here reads anything but a far-field.

   ## The gap

   The interesting version of the question lives in the medial superior olive — the first place in the ascending pathway where the two ears are compared. There is a far-field potential from phase-locked activity there, the neurophonic.

   The constraint comes from modelling work on that signal.

   {% include source.html key="goldwyn2014" %}

   Engagement here was with the released model rather than the paper's full text, so only what is checkable is stated: the published code contains **no sodium conductance and no axon** — only an h-current and a low-threshold potassium current — and it reproduces the far-field with no spike-generating machinery in it.

   Taken seriously rather than routed around. A matched far-field licenses *drive arrived at the nucleus*. It does not license *cells crossed threshold*. Those are different statements and a postsynaptic far-field cannot separate them. Most of the downstream biology of interest depends on the second.

   So the defensible claim is smaller: **can a designed acoustic protocol put a specified structure into a recorded evoked response, such that it tracks a parameter sweep and is distinguishable from a matched-energy plain tone?** A delivery-verification question. A methods result. Runnable.

   Except for one thing.

   ## The analog ceiling

   The board passes roughly 0.16 to 40 Hz. The signal needed sits at carrier rates in the hundreds of hertz.

   Not a limitation to work around. Reviewing the logs, **three consecutive workstreams were designed against an instrument whose bandwidth removes the target signal.** The question was fitted to the rig rather than the rig chosen for the question. A reasoning failure, not a hardware problem.

   Fatal rather than inconvenient because the filter is in silicon *ahead of the converter*. A 100 Hz fundamental has **zero harmonics inside the passband.** No dwell, no averaging, no clever choice of stimulus frequency recovers what was removed before digitisation. Hence the viability gate — *is the readout in the analog band at all* — before building anything.

   A partial escape: under dichotic drive the response envelope can modulate at the difference frequency even though no envelope exists acoustically at either ear. That appears spectrally as **sidebands around the carrier**, not as energy at the difference frequency — searching at the difference frequency is a null about a frequency the physics never predicted. Whether a single-channel Oz–Cz montage can recover sideband structure at all is open, and it is a hardware question.

   ## What I'd like

   Not a collaborator, co-author or funding. A competence I do not have.

   1. **A bandwidth reality check.** Is a sub-$100 board recoverable into a usable range with better front-end filtering and averaging, or is the ceiling the ceiling?

   2. **A montage correction.** Oz–Cz was chosen for alpha, then inherited by a question it does not suit. What would you use?

   3. **A benchmark, if there is appetite.** How far does consumer-grade EEG extend before it stops recovering what lab-grade systems get, and where is the boundary? Publishable with or without this programme attached, and I would run the tedious conditions.

   Neither read at full text. Mouse tissue readout, not scalp EEG. Precedent that the shape exists, nothing more.

   A bench test is queued that will probably close a large part of this programme — a sustained-response decay measurementtest is queued. If it fails, most of what was built downstreamon top of it was wasted effort. **A null is a result and it will be published.**

   ## Devil's advocate

   **A distraction from the bench.** The sustained-responsedecay test needs a headphone, an electrode and an afternoon. It has been the named next step for weeks. Writing an outreach postthis is a legible way not to not run it, and that risk is live.

   **The interesting claim and the defensible claim have come apart.** What is defensible is a delivery-verification methods result. What prompted the work was more ambitious. A hostile reader can grant the first and deny it bears on the second, and nothing here answers them.

   **The instrument figures above are weaker than their first version.** The median, the ratio and the attribution all changed on inspection. Anything else from this rig deserves the same recomputation before it is trusted.

   **A warm response does not solve the actual problem.** BetterA better instrument, correctedor montage —does neithernot getsgive tissue access andor pharmacological control. That is a different building, probably a different institution, possibly a different species.

   ## Correction — 2026-09-13

   This post originally stated a median in-band noise at 40 Hz of **238 nV across eight sessions**, a broadband overstatement of **20–90×**, that the spread showed **electrode preparation dominates**, and that the delivered stimulus envelope was **measured rather than assumed**. None held.

   The 238 nV came from an analysis script that took the upper-middle of six values; the median is 157 nV, over six captures. "20–90×" was a hard-coded literal in the same script's output; computed per capture it is 3.5–24×. The session records contradict the electrode-prep attribution. No stimulus-timing measurement exists. All four are corrected at source in the research records.

   A paragraph noting that beta-band and dichotic runs had no producible records was removed rather than kept as anecdote. The post was restructured to lead with the instrument characterisation.

   Commit 08f8d3b on GitHub →

4. v6 2026-09-12 07:54 posts: inline source cards, and every post cut to the point Joshua Anderton with Claude Opus 5 +187 −422 words

   status: published

   inline_sources: true

   tier: C

   ---

   I've been running a soloSolo research program forprogramme, about two months old. The honest next step is to have someone tell me mythe instrument is inadequate.

   This is the rig, what it has shown, and the gap. It's writtenWritten for people who validate portable EEG hardware for a living. The ask is at the bottom and it's is small.

   ## The rig

   A singleSingle-channel Olimex SHIELD-EKG/EMG on an Arduino. **Active Oz, reference Cz, ground mastoid** — a posterior montage chosen for alpha. 250 Hz sampling, 10-bit ADC at roughly

   7.9 µV per count, and a **~40 Hz analog low-pass**. Closed-back headphones. Python doing Welch PSD with FOOOF for aperiodic separation, and; Web Audio generating the stimulus, so the delivered envelope is measured rather than assumed.

   Microvolt EEG on a 10-bit converter is only a few counts. It's quantisationQuantisation-limited and low-SNR by construction, and I want to be precise about what it has and hasn't shown.

   ## Why this board, and what the alternatives cost

   ## The alternatives, and why this board

   I bought the cheap boardBought on hearsay — forum consensus that the expensive oneboard buys a quieter front end and convenience, not a capability you can'tcannot otherwise reach. **That's is not a measurement.** I've never seen aNo matched-conditions noise-floor comparison of the two has been seen here, and Inone didn'twas run one, because running one means owning both.

   Here's the actual landscapeLandscape, checked 2026-08-25:

   | | channels | price |

   | OpenBCI Galea | — | $60,504 |

   The gap between rows one and three is the whole question. Two failure modes Ithat can'tcannot rulebe ruled out from here: the noise-floor claim may hold for alpha and fail for anything smaller; or I optimised the wrong axis was optimised entirely, in which case **neither** Olimex board solves mythe problem and cheap-versus-expensive was never the interesting comparison.

   ## What theThe noise floor actually is

   Median in-band noise at 40 Hz over 60 s: **238 nV**, ranging 65 to 750 nV across eight sessions. An **11× spread on the same hardware in the same room** — electrode prep dominates everything the software does.

   One correction worth passing on, because it pointed the dangerous way. I'd been carrying aA broadband RMS figure had been carried as if it were the noise floor at the frequency of interest. Real EEG is drift-dominated, so broadband RMS **overstated 40 Hz noise by 20–90× in amplitude**. That error declares a runnable measurement unrunnable, which quietly kills experiments you should have attempted.

   **overstated 40 Hz noise by 20–90× in amplitude.** That error declares a runnable measurement unrunnable, which quietly kills experiments that should have been attempted.

   ## What I'vehas actuallybeen run

   Isochronic tones at the alpha resonant frequency, looking for a measurable aftereffect.

   IReported want to report this the wayas the data reads, rathernot thanas the way I remember itremembered. **Of six resonance profiles on file, five are `inconclusive`** — flagged by the app itself as drift-dominated, or with no alpha peak above the 1/f floor, or with a peak below the alpha band entirely. The one session marked `strong` sits at 10 Hz with 2.8 dB SNR. MyThe best recorded run gives a resonant frequency of 11 Hz against a peak of 9.77, a divergence of 1.23 Hz, and its own verdict is **`LOW QUALITY / no valid comparison`**.

   So: the chain demonstrably recovers alpha, and Ithat treat thatcounts as **rig validation and nothing else** — evidence that board, electrode, amplifier and analysis code together find a real, expected signal. It is aA calibration, not a finding. n = 1, unblinded, no sham for arousal or attention, effect measured *during* drive rather than after, and acute entrainment of an ongoing oscillation is among the least surprising results in the auditory literature.

   I also ran aA beta-band version of this and experimented with dichotic drive —experiments awere differentalso tone in each ear, 10 Hz apart, with an isochronic envelope on toprun. **I'm not reportingNo numbers are reported for either, because Ithe can'trecords cannot currently producebe the recordsproduced.** The beta app was retired and its results aren'tare not where I expected them. Until Ithe data can be put the data in front of you, those are anecdotes and I'd rather say so.

   ## The published precedent

   The nearest thing in the literature to the shape I'mbeing chasingchased is the MIT gamma work — 40 Hz sensory drive producing downstream cellular consequenceswork in mouse models {% include cite.html key="iaccarino2016" %}, later extended to combined auditory and visual stimulation {% include cite.html key="martorell2019" %}.

   {% include source.html key="iaccarino2016" %} {% include source.html key="martorell2019" %}

   I cite theseCited as **precedent that the shape exists**, not as support. INeither haven'twas read either at full text,. and fourFour differences bound how far they travel: mouse versus human, invasive tissue readout versus scalp EEG, 40 Hz gamma versus alpha, cortex and hippocampus versus a brainstem nucleus. **The readout difference is the binding one.** Those studies read the consequence in tissue. Nothing Iavailable can dohere reads anything but a far-field.

   ## The gap

   The interesting version of mythe question lives in the medial superior olive — the first place in the ascending pathway where the two ears are compared. There's is a far-field potential from phase-locked activity there, the neurophonic.

   The constraint comes from modelling work on that signal.

   {% include source.html key="goldwyn2014" %}

   TheEngagement constrainthere comes from the modelling work on that signal {% include cite.html key="goldwyn2014" %}. I've engagedwas with the released model rather than the paper's full text, so I'll stateonly what's is checkable is stated: the authors' published code contains **no sodium conductance and no axon** — only an h-current and a low-threshold potassium current — and it reproduces the far-field with no spike-generating machinery in it at all.

   I've taken that seriously rather than routing around it. A matched far-field licenses

   Taken seriously rather than routed around. A matched far-field licenses *drive arrived at the nucleus*. It does not license *cells crossed threshold*. Those are different statements and a postsynaptic far-field can'tcannot separate them. Most of the downstream biology Iof care aboutinterest depends on the second, and I can't get there with this instrument — or, as far as I can tell, with any non-invasive one.

   So the defensible claim is smaller: **can a designed acoustic protocol put a specified structure into a recorded evoked response, such that it tracks a parameter sweep and is distinguishable from a matched-energy plain tone?** A delivery-verification question. A methods result. Runnable.

   ## The analog ceiling

   The board passes roughly 0.16 to 40 Hz. The signal I needneeded sits at carrier rates in the hundreds of hertz.

   ThisNot is not "a limitation to work around." Reviewing my ownthe logs I found that, **three consecutive workstreams were designed against an instrument whose bandwidth removes the target signal.** The question was fitted to the rig rather than the rig chosen for the question. That's aA reasoning failure, not a hardware problem, and I'd rather name it than have someone else find it.

   It's fatalFatal rather than inconvenient because the filter is in silicon *ahead of the converter*. A 100 Hz fundamental has **zero harmonics inside the passband.**. No dwell, no averaging, no clever choice of stimulus frequency recovers what was removed before digitisation. IHence now run athe viability gate — *is the readout in the analog band at all* — before building anything, because I built an entire apparatus first and computed that gate afterwards.

   There's aA partial escape.: Underunder dichotic drive the response envelope can modulate at the difference frequency even though no envelope exists acoustically at either ear. That appears spectrally as **sidebands around the carrier**, not as energy at the difference frequency — searching at the difference frequency is a null about a frequency the physics never predicted. Whether a single-channel Oz–Cz montage can recover sideband structure at all is open, and it's is a hardware question.

   ## What I'd like

   Not a collaborator, co-author, or funding. A competence I don'tdo not have.

   1. **A bandwidth reality check.** Is a sub-$100 board recoverable into a usable range with better front-end filtering and averaging, or is the ceiling the ceiling?

   2. **A montage correction.** Oz–Cz was chosen for alpha, then inherited by a question it doesn'tdoes not suit. What would you use?

   3. **A benchmark, if there's is appetite.** How far does consumer-grade EEG extend before it stops recovering what lab-grade systems get, and where exactly is the boundary? That's publishablePublishable with or without mythis programprogramme attached, and I'd would run the tedious conditions.

   I have aA bench test is queued that will probably close a large part of mythis programprogramme — a sustained-response decay measurement. If it fails, most of what I'vewas built downstream was wasted effort. **A null is a result and Iit will publishbe itpublished.**

   If you have five minutes and an opinion about my montage, I'd like to hear it.

   ## Devil's advocate

   **It's aA distraction from the bench.** The sustained-response test needs a headphone, an electrode, and an afternoon. It's has been the named next step for weeks. Writing an outreach post is a legible way to not run it, and that risk is live.

   **The interesting claim and the defensible claim have come apart.** What Iis can defenddefensible is a delivery-verification methods result. What madeprompted methe startwork was more ambitious. A hostile reader can grant the first and deny it bears on the second, and nothing here answers them. That's the correct state to be in and I'd rather sit in it visibly.

   **Even aA warm response doesn'tdoes not solve mythe actual problem.** Better instrument, corrected montage — neither gets me tissue access and pharmacological control. That's is a different building, probably a different institution, and possibly a different species.

   Commit fb7ef7b on GitHub →

5. v5 2026-09-03 15:17 site: images on every post, bigger thumbnails, CTA moves into the sidebar Joshua Anderton with Claude Opus 5 +53 −0 words

   - goldwyn2014

   image: /assets/posts/eeg-recording-cap.jpg

   image_alt: >-

   A person wearing a multi-electrode EEG recording cap in profile, ringed

   electrodes and their lead wires covering the scalp.

   image_credit: Chris Hope

   image_license: CC BY 2.0

   image_license_url: https://creativecommons.org/licenses/by/2.0/

   image_source_url: https://commons.wikimedia.org/wiki/File:EEG_Recording_Cap.jpg

   ---

   Commit cb0df37 on GitHub →

6. v4 2026-09-02 20:46 posts: publish the cheap-EEG validation-gap post Joshua Anderton with Claude Opus 5 +1 −1 words

   project: audio-evoked-potentials

   status: draftpublished

   tier: C

   Commit 8861977 on GitHub →

7. v3 2026-08-31 15:57 site: adopt the monorepo design system, add two posts, publish the AI disclosure Joshua Anderton with Claude Opus 5 +0 −0 words

   ## The rig

   A single-channel Olimex SHIELD-EKG-/EMG on an Arduino. **Active Oz, reference Cz, ground mastoid** — a posterior montage chosen for alpha. 250 Hz sampling, 10-bit ADC at roughly

   7.9 µV per count, and a **~40 Hz analog low-pass**. Closed-back headphones. Python doing Welch PSD with FOOOF for aperiodic separation, and Web Audio generating the stimulus, so the delivered envelope is measured rather than assumed.

   |---|---|---|

   | Olimex SHIELD-EKG-/EMG *(what I have)* | 1 | **€19.95** |

   | Olimex EEG-SMT | 2 | €99.00 |

   Commit 993e5af on GitHub →

8. v2 2026-08-26 07:27 post: GENUS precedent, live pricing, montage, and a tightening pass Joshua Anderton with Claude Opus 5 +751 −1018 words

   description: >-

   Notes on anAn n-of-1 EEG rig, an analog ceiling that removes the signal I was designing

   designing for, and a question I can't answer alone.

   date: 2026-08-25 09:00:00 -0600

   citations:

   - iaccarino2016

   - martorell2019

   - goldwyn2014

   ---

   I've been running a solo research program for about two months,. and I've reached the point where theThe honest next step is to have someone tell me my instrument is inadequate.

   This post is a description of the rig, onewhat resultit has shown, and onethe gap. It's written for people who validate portable EEG hardware for a living — a category that includes at least one lab a fifteen-minute drive from where I'm sitting. If that's you, theThe ask is at the bottom and it's small.

   ## The setuprig

   A single-channel Olimex SHIELD-EKG-EMG on an Arduino, Oz–Cz montage, closed-back headphones, and a Python stack doing Welch PSD with FOOOF for aperiodic separation. Stimulus generation is Web Audio in the browser, which means the delivered envelope is measured rather than assumed — a distinction that turns out to matter, because hard gating of an amplitude-modulated tone injects a second harmonic that a raised-cosine ramp doesn't.

   A single-channel Olimex SHIELD-EKG-EMG on an Arduino. **Active Oz, reference Cz, ground mastoid** — a posterior montage chosen for alpha. 250 Hz sampling, 10-bit ADC at roughly

   7.9 µV per count, and a **~40 Hz analog low-pass**. Closed-back headphones. Python doing Welch PSD with FOOOF for aperiodic separation, and Web Audio generating the stimulus, so the delivered envelope is measured rather than assumed.

   That'sMicrovolt theEEG wholeon apparatusa 10-bit converter is only a few counts. It's aquantisation-limited hobbyistand riglow-SNR by any reasonable standardconstruction, and I want to be precise about what it has and hasn't shown.

   ## Why thethis EKG/EMG shieldboard, and notwhat the EEG-SMTalternatives cost

   This is the first decision I'd like checked, because I made it on hearsay.

   Olimex sells a purpose-built EEG board — the EEG-SMT — and I didn't buy it. It's roughly five times the price, and the argument I found in hobbyist EEG communities was that the SMT's advantage is essentially a quieter front end, which you can approach on the cheaper shield with careful grounding, shielded leads, good electrode prep, and a short cable run. The reasoning was that the expensive board buys convenience and a better noise floor out of the box, not a capability you can't otherwise reach.

   The epistemic status of that matters. **It's forum consensus, not a measurement.** I haven't seen a side-by-side noise-floor comparison of the two boards under matched conditions, and I didn't run one, because running one requires owning both — which defeats the purpose of the decision. So I bought the cheap board on an unverified claim and then built two months of work on top of it.

   That may well have been fine. But there are two failure modes I can't rule out from where I'm standing:

   - **The noise-floor claim may hold for alpha and fail for anything smaller.** A board adequate for a 10 Hz cortical rhythm at microvolt scale says nothing about a far-field brainstem potential two or three orders of magnitude down.

   - **I may have optimised the wrong axis entirely.** If analog bandwidth is the binding constraint — see below — then neither Olimex board solves my problem and the cheap-versus-expensive question was never the interesting one. Cheap and wrong is not better than expensive and wrong.

   If someone with a calibrated signal generator has ever put these two boards on a bench together, I'd like to see the numbers.

   I bought the cheap board on hearsay — forum consensus that the expensive one buys a quieter front end and convenience, not a capability you can't otherwise reach. **That's not a measurement.** I've never seen a matched-conditions noise-floor comparison of the two, and I didn't run one, because running one means owning both.

   Here's the actual landscape, checked 2026-08-25:

   | | channels | price |

   |---|---|---|

   | Olimex SHIELD-EKG-EMG *(what I have)* | 1 | **€19.95** |

   | Olimex EEG-SMT | 2 | €99.00 |

   | OpenBCI Cyton | 8 | $1,759 |

   | OpenBCI Cyton + Daisy | 16 | $3,518 |

   | OpenBCI Complete Ultracortex | 16 | $4,222 |

   | OpenBCI Galea | — | $60,504 |

   The gap between rows one and three is the whole question. Two failure modes I can't rule out from here: the noise-floor claim may hold for alpha and fail for anything smaller; or I optimised the wrong axis entirely, in which case **neither** Olimex board solves my problem and cheap-versus-expensive was never the interesting comparison.

   ## What the rig's noise floor actually is

   Since writing the above I've measured the part I can measure, across eight sessions on this rig.

   The headline is that **electrode prep dominates everything the software does.** Median in-band noise at 40 Hz over a60 60-second window iss: **238 nV**, with a session-to-session range ofranging 65 to 750 nV —across aneight sessions. An **11× spread on the same hardware in the same room.** No— processingelectrode choiceprep Idominates have made moveseverything the result as much as that spreadsoftware does.

   One correction is worth passing on, because it pointed the dangerous way. I had'd been carrying a broadband RMS figure as thoughif it were the noise floor at the frequency of interest. Real EEG is drift-dominated, so a broadband RMS **overstated the noise at 40 Hz noise by 20–90× in amplitude** — thousands of times over in required dwell. That error declares a runnable measurement unrunnable, which is the direction that quietly kills experiments you should have attempted.

   ## TheWhat oneI've resultactually run

   Isochronic auditory drive at individual alpha frequency (~9.7 Hz for me) produces roughly a **22% rise in alpha-band power at IAF** during stimulation, reproducibly, on-rig.

   Every caveat you're already forming is correct. **n = 1.** No blinding. No sham designed to control for arousal or attention. Single channel, non-standard montage, and Oz–Cz is not what anyone would choose for this if they had a choice. The effect is acute — measured *during* drive, not after — and acute entrainment of an ongoing oscillation is one of the least surprising things in the auditory literature.

   Isochronic tones at the alpha resonant frequency, looking for a measurable aftereffect.

   I want to report this the way the data reads rather than the way I remember it. **Of six resonance profiles on file, five are `inconclusive`** — flagged by the app itself as drift-dominated, or with no alpha peak above the 1/f floor, or with a peak below the alpha band entirely. The one session marked `strong` sits at 10 Hz with 2.8 dB SNR. My best recorded run gives a resonant frequency of 11 Hz against a peak of 9.77, a divergence of 1.23 Hz, and its own verdict is **`LOW QUALITY / no valid comparison`**.

   So: the chain demonstrably recovers alpha, and I treat thisthat as **rig validation and nothing else**: — evidence that the board, the electrode, the amplifier chain and my analysis code together can recoverfind a real, expected signal. It's is a calibration, not a finding. n = 1, unblinded, no sham for arousal or attention, effect measured *during* drive rather than after, and it'sacute filedentrainment atof an ongoing oscillation is among the lowestleast tiersurprising thisresults projectin hasthe —auditory a first-person unblinded observation, which no amount of repetition can promoteliterature.

   One further honesty note. When I went to record that result formally, I found it had only ever existed as a number quoted inside a design document — the sessions were never logged at the granularity that would let me recover exact dates or per-session figures. So read the 22% as approximate, and read its registration as retroactive, which is weaker than pre-registered.

   I also ran a beta-band version of this and experimented with dichotic drive — a different tone in each ear, 10 Hz apart, with an isochronic envelope on top. **I'm not reporting numbers for either, because I can't currently produce the records.** The beta app was retired and its results aren't where I expected them. Until I can put the data in front of you, those are anecdotes and I'd rather say so.

   ## The gap,published stated plainlyprecedent

   The nearest thing in the literature to the shape I'm chasing is the MIT gamma work — 40 Hz sensory drive producing downstream cellular consequences in mouse models {% include cite.html key="iaccarino2016" %}, later extended to combined auditory and visual stimulation {% include cite.html key="martorell2019" %}.

   I cite these as **precedent that the shape exists**, not as support. I haven't read either at full text, and four differences bound how far they travel: mouse versus human, invasive tissue readout versus scalp EEG, 40 Hz gamma versus alpha, cortex and hippocampus versus a brainstem nucleus. **The readout difference is the binding one.** Those studies read the consequence in tissue. Nothing I can do reads anything but a far-field.

   ## The gap

   The interesting version of my question lives in the auditory brainstem, in the medial superior olive — the first place in the ascending pathway where signals from the two ears are compared. There's a far-field potential associated withfrom phase-locked activity there, called the neurophonic.

   The constraint I've had to absorb comes from the modelling work on that potentialsignal {% include cite.html key="goldwyn2014" %}. I want to be careful about how I state this, because I have've engaged with the released model rather than with the paper's full text., Whatso I'll can verify directly isstate what's incheckable: the authors' own published code: contains **the model contains no sodium conductance and no axon at all** — only aan hyperpolarisationh-activated current and a low-threshold potassium current. It— and it reproduces the far-field signalwith without anyno spike-generating machinery in it at all.

   I've taken the implicationthat seriously rather than routing around it. A matched far-field licenses the claim *phase-locked drive arrived at the nucleus*. It does not license

   *drive arrived at the nucleus*. It does not license *cells crossed threshold*. Those are different statements, and a postsynaptic far-field cannotcan't distinguishseparate them. MuchMost of the interesting downstream biology I care about depends on the second one, and I can't get there with thethis instrument I have — or, as far as I can determinetell, with any non-invasive instrumentone.

   So the defensible claim I'm left with is smaller and more defensible: **can a designed acoustic protocol put a specified structure into a recorded evoked response, such that the structureit tracks a designed parameter sweep and is distinguishable from a matched-energy plain tone?** That's aA delivery-verification question. It's aA methods result. It's also runnableRunnable.

   Except for one thing.

   ## The analog ceiling

   The Olimex board passes roughly 0.16 to 40 Hz. The signal I need to characterise sits at carrier rates in the hundreds of hertz.

   I want to be blunt about what that means, because I've caught myself doing the softer version. It is not "a limitation to work around." Reviewing my own logs, I found that

   This is not "a limitation to work around." Reviewing my own logs I found that **three consecutive workstreams were designed against an instrument whose bandwidth removes the target signal.** The question was being fitted to the rig rather than the rig chosen for the question. That's a reasoning failure, not a hardware problem, it's a reasoning failure, and I'd rather name it than have someone else find it.

   The reason itIt's fatal rather than inconvenient is thatbecause the filter is in silicon *ahead of the converter*. A 100 Hz response fundamental has **zero harmonics inside the passband**. No amount of dwell, no averaging, orno clever choice of stimulus frequency recovers something thatwhat was removed before the signal was digitiseddigitisation. I now run a viability gate — *is the readout in the analog band at all* — before building anything, because I built an entire apparatus first and computed that gate afterwards.

   There's a partial escape. Under dichotic drive — a slightly different pure tone to each ear — the envelope of the response can modulate at the difference frequency even though no envelope exists acoustically at either ear. That structure appears spectrally as

   There's a partial escape. Under dichotic drive the response envelope can modulate at the difference frequency even though no envelope exists acoustically at either ear. That appears spectrally as **sidebands around the carrier**, not as energy at the difference frequency itself.— Searchingsearching at the difference frequency is a null about a frequency the physics never predicted. Whether a single-channel Oz–Cz montage can recover sideband structure at all is genuinely open, and it's a hardware question, not a theory question.

   ## What I'd like

   I'm not looking forNot a collaborator, a co-author, or funding. I'm looking for aA competence I don't have.

   1. **A bandwidth reality check.** Is a sub-$100 board recoverable into a usable range with better front-end filtering and averaging, or is the ceiling the ceiling? I suspect I know, but I've been wrong about hardware before and I'd rather be told than discover it after another month.

   2. **A montage correction.** Oz–Cz was chosen for alpha, then inherited by a question it doesn't suit. What would you use?

   3. **A benchmark, if there's appetite.** There's a real methodological question here: howHow far does consumer-grade EEG extend before it stops recovering signalswhat lab-grade systems get, and where exactly is the boundary? That's publishable with or without my program attached, and I'd happily run the tedious comparison conditions.

   I have a bench test queued that will probably close a large part of my program — a sustained-response decay measurement, testing whether the candidate envelope holds above 50% of initial amplitude. If it fails, most of what I've built downstream was wasted effort, and I'd rather find out this month than next year. **A null is a result and I will publish it.**

   If you have five minutes and an opinion about my montage, I'd like to hear it.

   ## Devil's advocate

   Three objections I'd rather raise than have raised.

   **It's a distraction from the bench.** The sustained-response test needs a headphone, an electrode, and an afternoon. It has's been the named next step for weeks. Writing an outreach post is a legible, satisfying way to not run it, and that risk is live.

   **The interesting claim and the defensible claim have come apart.** What I can defend is a delivery-verification methods result. What made me start was considerably more ambitious. A hostile reader can grant the first entirely and deny it bears on the second, and nothing abovehere answers them. I think thatThat's the correct state to be in, and I'd rather sit in it visibly.

   **Even a warm response doesn't solve my actual problem.** EEGBetter methodologyinstrument, expertisecorrected montage — neither gets me a better instrument and a corrected montage. It doesn't get me what the ambitious half of the question requires, which is tissue access and pharmacological control. —That's a different building, probably a different institution, and possibly a different species.

   Commit d669e46 on GitHub →

9. v1 2026-08-26 06:48 draft: can a $70 board tell me I'm wrong? Joshua Anderton with Claude Opus 5 first version

   v1: the original version, as first committed. Read it on GitHub →

   Commit a079339 on GitHub →