The thing I built is not a simulator. A simulator implies you can turn a crank and get behaviour out the other end. What this does is narrower and, I think, more useful: for a given stimulus it answers how much arrives at a receptor, in what format that receptor passes it on, and where it goes — and it refuses to answer when the number underneath does not exist. Calling it a screening tool is more honest than calling it a simulator, and the difference matters, because the most common output is a refusal.

The design rule: every number carries where it came from

Each constant in the tool is tagged with its provenance — measured in this species, measured in a surrogate species, derived from physics, or assumed. That sounds like bookkeeping. It turned out to be the part that did the work.

Over the course of building it, the tags caught six errors that had already made it into the model, every one of them from reading a summary instead of a source. Two examples, because they are the kind of thing that is invisible without the discipline:

• A trap statistic entered the model as “entry success rises from 16% to 59%.” Those are two different measurements — one is the fraction of attempts that succeed¹, the other is a percentage increase in catch from a modification². The second does not continue the first, and the best modification was not the one being quoted.
• A receptor threshold was carried in the model as a single value when the underlying study reports a range³. The single number was the low end of it.

The rule that generates the rest: an abstract is not a paper. A source read only at abstract can be cited as existing, but nothing may be said about what it measured. That constraint is load-bearing in what follows, and I have marked where it bites.

The mechanical senses — the part that is genuinely measured

The strongest ground is mechanoreception, because there is a real transfer function for this species. Recordings from the mouthpart setae give displacement thresholds in micrometres, and — more useful — the coding: spike count scales with displacement amplitude, interspike interval with velocity, and roughly half the cells are directionally sensitive⁴.

That last detail is what made the first version of the tool wrong. Those setae are tactile receptors for handling prey, not distance detectors, and I had been scoring far-field stimuli against them. The tool now raises a scope flag on any far-field verdict from that row, because the answer it would give is meaningless.

The receptor that is for water motion is a different one: sensory hairs on the chelae, studied in crayfish⁵. A joint organ in the walking leg has been described in green crab itself⁶, the behavioural response of intact animals to substrate vibration has been tested directly⁷, and decapod acoustic detection has been reviewed as a field⁸. An evoked-potential study in the American lobster asks which organ is actually responsible⁹, and a consolidated review of green crab sensory biology exists¹⁰.

Where the abstract-only rule bites, and it bites here: of the seven sources in this section I have read two at full text. For the other five I can tell you they exist and what question they asked — not what they found. So the surrogate threshold the tool actually uses for the water-motion channel is a number I am not in a position to attribute, and the tool carries it tagged as a surrogate for exactly that reason.

The physics that decides most of it

Two pieces of arithmetic ended up governing nearly every conclusion.

Near-field falloff. A small oscillating body in water produces a field that falls off as 1/r³ if it oscillates back and forth, or 1/r² if it changes volume. The cube law is brutal: a thousand times more source amplitude buys ten times the range. Working through plausible prey sources, ordinary prey turns out to be a centimetre-scale cue; only a large struggling animal or another crab reaches decimetres.

Displacement is not pressure. Particle displacement equals velocity divided by 2πf, so for a given sound pressure a kilohertz signal carries far less displacement than a hundred-hertz one. High-frequency sound is displacement-poor, and crabs detect displacement.

Those two together explain why the acoustic route keeps failing, and they are also why the chemical route keeps winning: a plume falls off as 1/r and the olfactory threshold is low.

What prey actually emit

Having built the receiver side, the obvious next question is what is available to receive.

The reference dataset for benthic invertebrate sound gives calibrated source levels for scallops, limpets, urchins and crustaceans¹¹. Converting each to particle displacement and solving for the crab’s threshold, none of it reaches a crab at more than millimetres — and every sound in that dataset peaks between 5 and 49 kHz, while the analysis band was filtered from 2 kHz upward. The corpus was built for hydrophone monitoring, not for prey detection, and the crab’s band was removed before analysis. A null there is unavailable, not absent.

The cue that does look promising is not a sound at all. Porewater pressure signals generated by infaunal activity have been recorded directly in sediment¹², and the irrigation timing of three tellinid bivalves — including a Pacific Northwest clam that is real green crab prey — has been measured in detail¹³. That is a slow hydraulic signal, not an acoustic one, and it is the only prey cue in this review with a field detection range measured on relevant species.

Chemistry, which out-ranges everything

Modelled as a plume, chemical detection spans roughly three orders of magnitude more distance than any mechanical cue. The receptor side has two anatomically separate systems doing different jobs — olfaction on the antennules, and contact chemoreception on the dactyls, where the gustatory organs of this species have been characterised electrophysiologically³. I have that paper at abstract only, so the threshold separation between the two — which is what the tool actually runs on — is stated in the model as a sourced parameter and is not stated here.

The tool refuses one thing here, deliberately. A chemical stimulus is not a scalar. You cannot hand it a single “concentration” and get a receptor response, because mixtures cross-adapt and receptors encode a ratio against background rather than an absolute. Single-compound dose-response is enabled; mixtures raise an exception.

The female sex pheromone in this species has been the subject of an identification study¹⁴, and later work using it produced the single most interesting behavioural result in this review: males presented with a pheromone-treated dummy female showed a mating response 87% of the time under ambient conditions and 40% under ship-noise playback, a significant drop — while the time taken to respond to the pheromone did not change significantly¹⁵. Noise interfered with completing the behaviour, not with detecting the signal. That dissociation is worth more than either number alone.

Electric and magnetic — where the tool says no

These are the two channels the tool refuses outright, and the refusals are the highest-confidence outputs in it.

Electric. No electroreceptor has been identified in any crustacean, and none has been sought in this species. That is not a modelling gap; it is the state of the field. The physics side is unambiguous: prey bioelectric fields were measured at microvolt scale with the electrode less than a millimetre from the tissue, and the authors state plainly that the voltage from invertebrates was too weak to record away from the source¹⁶. A behavioural sensitivity figure exists for a freshwater crayfish¹⁷ — a surrogate, from a different order, in fresh water. The tool will quote named surrogates and will not produce a crab number.

Magnetic. There is a directly relevant result: juvenile green crabs were exposed to static magnetic fields at strengths relevant to submarine power cables, and females spent substantially more time in the exposed zones while males showed no consistent preference¹⁸. That is a real, sex-specific response in this species — and it is a static-field spatial preference, not a compass, and not a receptor. Nobody has located the transducer.

And there is a confound worth stating plainly, because it is the obvious one. Helmholtz coils dissipate heat, and crabs are thermotactic. Working the published coil geometry through a thermal model gives, at the strongest field over the exposure period, roughly 20 W dissipated, a 0.58 K gradient across the tank and a 0.30 K bulk rise — small, but not obviously below what an animal can detect, and I could not find a behavioural thermal-discrimination threshold for this species to compare it against. So the model returns a number and refuses a verdict.

One thing does argue against the thermal explanation, weakly: ohmic heating scales with the square of the current, so the confound is about ten times larger at the strongest field than at the intermediate one — while the reported attraction was strongest at the intermediate field. The confound and the effect run in opposite directions. A thermocouple in the tank would settle it, and that is the cheapest experiment anywhere in this review.

The usual comparison is the spiny lobster, which is the standard invertebrate magnetoreception system¹⁹²⁰²¹. Here the abstract-only rule bites hardest: I have read those three at abstract, so I will not characterise their results. What I will say is a statement about the shape of the literature rather than its contents — the receptor cells themselves have not been located and characterised in that animal either. Mapping a mechanism from one order of crustaceans onto another, across more than two hundred million years of divergence, when the source species’ anatomy is itself unresolved, produces a hypothesis wearing borrowed credibility. The tool returns unavailable and says why.

How the native crabs compare

Comparing green crab to the Pacific Northwest natives was the point at which the exercise became most informative, because the comparison cannot be run.

Across eight receptor channels and five species, the green crab has measured values in five channels and a described-but-unmeasured organ in two. For Dungeness crab, red rock crab, graceful rock crab and yellow shore crab, I found two measured thresholds, both in Dungeness. One establishes a behavioural detection threshold for a prey extract²⁴; the other, behavioural salinity detection measured by antennular flicking²⁵. I have both at abstract only, so those are descriptions of what was measured, not of what was found — the model carries the values with the same tag. (A third record sits off this matrix entirely: dactyl chemo- and mechanoreceptor recordings in the kelp crab, a species not among the five.)

Two filled cells out of thirty-two, and the comparison still cannot be run — but for a different and much smaller reason than an empty literature. Those are behavioural thresholds, in grams per litre and parts per thousand. Every green crab value is single-unit electrophysiology, in micrometres or molar. They are not the same quantity, so putting them in one row would be a category error rather than a comparison.

That is the state of the literature, not a gap in the modelling. Green crab is a classic laboratory animal; the natives are commercially and ecologically important animals whose sensory physiology has largely not been measured.

What can be compared is everything except the senses — and there the differences are substantial and well documented²²: salinity tolerance, depth and zonation, body size, moult timing, and tolerance of emersion and hypoxia. Subadult Dungeness use the littoral zone in ways that matter for any comparison drawn from trap data²³, red rock crab emersion physiology has been measured in situ²⁶, settlement and emigration behaviour has been studied in green crab²⁷, predatory capability has been compared directly between green crab and red rock crab on a native clam²⁸, and the effect of a recovering apex predator on green crab abundance has been examined²⁹.

I have read the first of those at full text and the rest at abstract, which is why this paragraph names questions rather than answers.

The genetics, and what is actually there

I went looking for a genetic basis for the green crab’s reproductive advantage — some identified difference that would explain why it out-reproduces the natives. There isn’t one, and the absence is informative.

The adaptive-genomics literature for this species is real and substantial. What it is about is temperature: cold tolerance, thermal adaptation, a chromosomal inversion, and the relationship between genotype and thermal plasticity³⁰³¹³². I have read these at abstract only and will not characterise their findings. But the observation that the surveyed literature concerns thermal physiology and not reproduction or sensory biology is an observation about the literature, and I can make it.

So the honest account of why green crab out-reproduce the natives is life-history and tolerance, not a gene: a longer breeding season, earlier maturity at smaller size, two broods a year where conditions allow, and sperm storage such that one fertilisation can serve more than one clutch²². Note also that reproducing more is not the same as mating more — with sperm storage, the two can move in opposite directions.

What the tool is for

The most-used outputs are refusals: no electric readout, no magnetic readout, no mixture readout, no cross-species comparison, no sex difference in transduction. Sex enters the model in exactly one place — body size — because no measured threshold in this species differs by sex, and the tool will not invent one.

That is a strange thing to build on purpose, and it is the part I would defend. The value is not in the numbers it produces. It is that the numbers cannot drift: every constant is pinned by a test that fails if the figure is misquoted, and in one working session those tests caught four errors that had entered the model from secondary summaries of paywalled sources — which is exactly how this literature is normally read.

The most instructive failure, though, was mine and not the literature’s. The claim above — that no native species had a single measured sensory channel — was a strong negative drawn from a shallow search, and it was wrong; a deeper search turned up the two Dungeness thresholds immediately. A test had been pinning the wrong number in place, with a comment beside it admitting the value was known to be stale. That is worse than an unchecked number, because it looks checked.

And correcting it nearly introduced a new error. The function answering which channels are comparable across species tested only whether a cell was marked measured. Filling the two Dungeness cells would have made two channels report as comparable when the underlying quantities are not the same thing. It compares units now. A test that counts statuses will happily certify a category error, and the correction is what exposed it.

What would make it wrong: if a measured audiogram for Carcinus maenas, Dungeness and red rock crab on one rig showed the receptor bands to be substantially different from the surrogate values used here, most of the mechanical conclusions would need redoing. That measurement does not exist, and it is the single experiment that would change the most.

Corrections

2026-09-06. Three changes, none of which erase anything above.

• The native-comparison section originally said I had found “two measured chemosensory thresholds in a single species and nothing else.” The two are now cited by key²⁴²⁵, one of them is a salinity threshold rather than a prey-odour one, and “nothing else” was an overreach — a third native record exists off the matrix. Both are held at abstract only, so the post names what they measured and not what they measured it to be. The count of filled cells is unchanged at two of thirty-two, and the comparison remains un-runnable.
• The “What the tool is for” section gains the account of how that error survived a test, and of the second defect that correcting it exposed.
• The correction has been applied to the model itself, not only to this post.

The starting observation is unremarkable once stated: an ECG is not only a record of a heart contracting. The published literature has read identity, sex, age and health out of the same trace. If that is true of one signal, the obvious question is what every biological signal carries, and whether the same structural features carry it each time.

Answering that properly starts with a boring inventory.

The signals catalogued so far

Six, named by the tissue they come from rather than the application they serve. These are representative textbook ranges, not hard bounds.

Writing that table was the first useful thing, and not for the reason I expected. Laid out together, the six are separated by properties that have nothing to do with physiology: how periodic they are, how sharp their events are, where their spectral weight sits, how much DC they carry. You can tell them apart without knowing what tissue you are looking at. That is either an interesting fact about biological signals or an artefact of how they are all recorded, and I genuinely do not know which yet.

The axes, and where each one comes from

The second list is the map: what kind of information is a biological wave known to carry? Each row below is an entry point to a literature, not a claim of mine.

Identity. Attributing a trace to a specific person is a mature field for ECG ¹, and there is a parallel line of work on individuating signatures in EEG ².

Sex and age. Both have been read from short clinical recordings ³, and the discrepancy between an age estimated from the signal and the person’s actual age has itself been studied as a marker ⁴.

Autonomic and affective state. Heart-rate variability is the standard non-invasive window onto autonomic balance, and electrodermal activity indexes a sympathetic pathway directly.

Intention. Muscle activity decodes motor intent; cortical activity carries potentials that precede movement rather than follow it.

Health. The axis with the most literature behind it, and the only one this bench has measured anything on directly.

I want to be careful here, because this is exactly the point where a survey turns into an overclaim: I have not read any of those papers at full text. They are cited as the entry point to their topic. What I can say is that these axes exist and are established; what I cannot say from this page is what any individual study found.

The part I did not expect

Building the list changed what I thought the project was.

I had been treating this as signal processing — a question about extracting a feature from a trace. Laying the axes out next to each other makes it look like a taxonomy problem instead. Identity, sex and age are properties that barely change. State and intention change minute to minute. Health drifts. Those are not the same kind of quantity, and lumping them together as “information in the signal” hides the most important distinction between them.

That reframing is the actual output of the exercise, and it came from making a table, not from running anything.

Pointing it at waves with nothing alive in them

Here is the part that keeps the whole thing honest, and it is the reason I would defend this method to anyone building something similar.

If you build a reader that finds rich structure in biological signals, you have to ask whether it finds rich structure in everything. So the same reader gets pointed at waves with no organism anywhere in them:

• Tide gauges — two stations, 61 days. A blind read recovered the principal lunar semidiurnal period and correctly separated the moon’s contribution from the weather’s.
• Solar flux against cosmic-ray flux — 14.6 years. It recovered the known anticorrelation, but only when given the full cycle. That makes it a result about how much data you need, not about the sun.
• Seismic, infrasound, ocean sound and the geomagnetic field — it recovered the secondary microseism and the daily solar-quiet variation, blind.

The standing rule on that log is the sentence I would put on the wall:

A high structure score on a signal with no organism in it is a caution about the instrument, not a finding about nature.

Reading tides well is partly bad news. It means the structure the reader is finding might be a property of the reader rather than of life. Every one of those runs was a test of the instrument that could have embarrassed it, and two of them did — one exposed silent bugs at low sample rates, another showed a standard statistical null to be invalid on deterministic signals.

Sound, light, and water

The obvious next question is whether the same characteristics carry information in non-biological waves, and here I want to be honest about the state of it: this is mapped, not tested.

The structural properties in question — how loud, how bright, how periodic, how consistent the repeating unit is, whether a slow rhythm modulates a fast one — all have natural readings in sound and in light. Loudness and timbre in acoustics; brightness and colour in optics; amplitude and morphology in a bioelectric trace. Water gave the cleanest test so far precisely because a tide is so well characterised that the reader had nowhere to hide.

But a correspondence you can write down is not a correspondence you have shown. Applying these characteristics across modalities is, until it is tested the same way the cardiac work was, a hypothesis carried by analogy. I would rather say that plainly than let a tidy table imply otherwise.

What would make this wrong

If the structural characteristics that separate biological signals turn out to separate any filtered time series equally well, then the shared-grammar idea is an artefact of everyone using similar analysis pipelines, and this survey is describing the pipeline rather than describing life. That is a real possibility, it is the reason the non-biological waves get read at all, and nothing on this page rules it out.

The short version: software buys you enormously more than I expected, and none of it is sensitivity. This is a walk through what was actually tried, what the other implementations in this space do, and where the line sits.

The mentality

The instinct when you have inadequate hardware is to look for the clever trick — the processing step that recovers what the sensor missed. That instinct is mostly wrong, and it took a while to internalise why.

What software can do, it turns out, is a long list:

• Verify rather than assume. Every parameter you did not measure is a parameter you are guessing at.
• Know its own noise, in the band that matters rather than across the whole spectrum.
• Refuse to believe itself — surrogate nulls, held-out thresholds, injection tests that establish what the pipeline would have detected.
• Keep everything on one clock, so that motion can be separated from field.
• Report its own gaps instead of quietly interpolating them.

What software cannot do is manufacture information that never entered the converter. If the quantisation step is larger than the thing you want to see, no amount of averaging, whitening or filtering invents it. That is the whole lesson and it took building the thing to learn it properly.

The other implementations, and what each one is for

I did not build this in a vacuum, and the most useful hours were spent reading what other people had already done. Four are worth documenting, because each is built for a different purpose and each is better than mine at its own job.

NOAA’s CrowdMag collects smartphone magnetometer readings as a distributed geomagnetic survey. Its defining property is that it validates every contribution against a field model — it knows whether a reading is correct, in absolute terms. Mine does not, and has never checked.

phyphox and Physics Toolbox are physics-education tools. Their defining property is that they expose the raw and the calibrated channel side by side, switchable in two taps. That distinction matters more than it sounds and I will come back to it.

The NASA/Odenwald citizen-science work characterises what a phone magnetometer can do for space-weather monitoring ², and is the closest published reference point for the noise performance of this class of sensor.

ADVIO is a visual-inertial odometry research dataset — not a magnetometry project at all, but it ships real multi-sensor phone recordings under an open licence, which makes it a cross-device control for anyone testing whether an effect is a property of one handset or of the platform.

How I use them. CrowdMag is the model for an absolute-accuracy check I have not yet built. phyphox is the reason I know the dual-channel decision is a defect and not a preference. ADVIO is the corpus that lets a finding be checked on hardware I do not own. None of that required permission; all of it required reading.

What was actually tried

Measure the delivered sample rate, per recording. Not the requested rate — the delivered one. It is not the same number, the difference is systematic, and every frequency-domain result is wrong by that factor if you assume instead of check. None of the implementations above verifies this per record.

Measure the noise floor in-band, not broadband. This one produced the most useful correction of the whole project, in the opposite direction from the one you would fear: a broadband figure overstated the noise at the frequency of interest by a large factor, because real recordings are drift-dominated. Carrying the wrong number around declares runnable experiments unrunnable. An instrument that makes you too pessimistic kills experiments silently, and nothing in the output tells you it happened.

Build the nulls before the result. Four families of surrogate, thresholds taken as a maximum statistic over the entire search rather than per-cell, and a refusal to quote a false-alarm bound above a stated ceiling. The comparators run quality control against a model; none of them runs a null. This is the single biggest difference and it is entirely software.

Validate by injection. Put a known signal into the real path and measure what fraction the pipeline recovers. Without this you have no idea whether a null result means absence or means your analysis would have missed it anyway.

One monotonic timebase across every sensor stream. Magnetometer, accelerometer, gyroscope, attitude — all on one clock. Almost every useful discriminator for separating “the phone moved” from “the field changed” depends on this, and it is impossible to retrofit.

Where it is worse, including one real defect

Being honest about this is most of the value of having done the comparison.

The calibrated channel is not logged, and that is a genuine defect. Recording the raw channel as primary is right — calibration is a filter with unpublished behaviour, and you cannot undo it. But the calibrated channel alongside it is a free diagnostic, phyphox does it, and its absence directly caused a multi-day misattribution: a large static field offset was read as an external accessory when it was internal to the device. A diagnostic channel I chose not to record would have shown that immediately. It is the cheapest fix on my list and it needs an app build.

No absolute accuracy validation. CrowdMag would have caught the same thing a different way.

One operating system, and a private archive. The cross-device control leans on other people’s corpora because mine is not shareable.

Dropped samples across long recordings, reported rather than hidden — and the “zero dropped gaps” figure that circulated in my own notes came from a single short run and does not survive contact with hours of recording.

The wall

The target for reading cardiac activity magnetically is described in the review literature ¹, and the gap between it and a phone is not a factor of a few. It is orders of magnitude, and it is a property of the sensor.

That is the part worth stating plainly, because it is where the mentality I started with runs out. Every item in the “what software can do” list above is real, and I would defend all of it as worth building. Not one of those things moves the floor. They tell you where the floor is, they stop you fooling yourself about what is above it, and they make a null result mean something — which is a great deal, and is not the same as sensitivity.

The honest summary: for what it is actually for — searching for weak periodic structure with real nulls — this is a more careful instrument than the comparators, and the gap is not close. For what the comparators are built for, it is worse, and one of those gaps is a defect rather than a trade-off.

What would change my mind

If someone demonstrates that a stock phone magnetometer resolves a signal at the amplitude the cardiac literature describes, under a protocol with real nulls and an injection test, then the sensitivity wall I am describing is not where I think it is and this post is wrong. I would want to see the null machinery before the result, which is the same standard I am holding myself to.

What this post is, exactly. Most of it describes phenomena at the level where they are settled and uncontroversial, and most of the sources below are cited as the entry point to their topic — not read here at full text, so nothing on this page reports what those papers measured or concluded. Where I say something is known, I mean textbook-known, not “this study showed.”

One source is different. The peafowl work has been read at full text, so that section does report what it found, and the post is tiered B on account of it rather than A. The reference list marks which is which; a badge on an entry reading full text read is the difference between “this paper exists” and “this paper found.”

Reading a field another body makes

The oldest example is the one everybody has heard of. Sharks, rays and skates detect the standing bioelectric field that surrounds another animal, using jelly-filled canals in the head called the ampullae of Lorenzini ¹. Seawater helps enormously here: it conducts, so a body’s field propagates through it and stays readable at a distance.

The two examples that impressed me more happen in air, which is much harder, because air does not conduct and there is no comparable field to swim through.

Bumblebees detect the electric field around a flower ². Spiders detect the atmospheric potential gradient — the standing voltage difference between the ground and the sky — and use it as a cue for ballooning ⁴.

Both solve the air problem the same way, and this is the part worth taking: neither uses anything like an electrode. They use hair ³. A charged filament sitting in an electric field experiences a force and moves, and both animals already had an organ exquisitely good at detecting the movement of hairs, because that is how they sense air currents and vibration. The field sense is a mechanical sense wearing a different hat.

For anyone building an instrument, that reframes the problem. A field detector does not have to be a voltage-measuring device in contact with something. It can be anything charged, compliant, and watched closely enough.

Reading a field you make yourself

Everything above is passive: the animal reads a field something else produced. There is a second mode, and it is the one an instrument builder should look at hardest, because it is the one where you supply the signal.

Weakly electric fish emit a discharge from a dedicated organ and read how nearby objects perturb it. That is active sensing — the same idea as radar or sonar, arrived at independently, in water.

What makes it worth studying is not that it works. It is what it costs. Four things have to be present, and each is separately necessary:

1. An organ that produces a discrete, command-triggered discharge.
2. A receptor class tuned to that discharge, distinct from the receptors the same fish uses to sense other animals passively.
3. A cerebellum-like structure that learns a negative image of the animal’s own discharge and subtracts it, through a plasticity rule that depends on the order in which the two inputs arrive ⁵.
4. A corollary discharge — a copy of the motor command that fired the organ — supplying the timing the negative image is built against ⁶.

Item 4 is the one I keep coming back to. The cancellation is not computed from the recorded signal. It is timed from the command that caused it, before the signal arrives. An instrument that emits and then tries to subtract its own contribution by measuring it is solving a harder problem than the fish is, and the fish has had a long time to look for an easier route.

The other half of the lesson is a warning about copying it. This works in water because water conducts, so the fish drives a current through the medium and objects perturb the current density at its skin. Air does not conduct. The nearest thing in air is capacitive — a nearby object changes how much charge sits where — which is a different quantity, with a different distance dependence, and it responds to a different property of the object. The intuition transfers; the arithmetic does not.

The spider web — and the story that is wrong

There is a well-travelled claim that a spider web’s geometry does something clever with light — that the web amplifies a fly’s photons, or exploits some optical property of the spiral. I went looking for it because it is a good story.

It is not what happens. What actually happens is electrostatic ⁷. Insects in flight carry charge. A web does not. When a charged insect passes close, the silk is pulled toward it — the threads physically deform in its direction, which makes contact more likely than the geometry alone would predict.

That is a better story than the optical one, for a reason worth stating plainly: there is no sensor anywhere in it. No receptor, no nerve, no organism doing anything. It is a passive dielectric structure whose shape converts a static charge difference into mechanical motion. Everything else in this post is an animal reading a field. This is geometry alone doing the work.

I mention the correction rather than quietly writing the right version because the two ideas get conflated constantly, and only one of them has a source behind it.

Signals that travel

Distance is the constraint that kills most sensing ideas, so it is worth knowing what biology manages.

In marine sediment there are filamentous bacteria that move electrons along their length, through conductive structures running the length of the filament ⁸. The distances involved are centimetres — which sounds modest until you hold it against the size of a cell, where it is enormous.

Plants do something different and just as long-ranged: a wound at one site produces a travelling calcium wave that reaches undamaged tissue elsewhere in the organism ⁹.

The reason to keep these two apart is that they are not the same kind of transport. Most of what I reason about on this bench is ionic — the action potential, charge moving across a membrane. Cable bacteria are not doing that. They are moving electrons, which is a different mechanism with different limits, and it is a useful reminder that “biological conduction” is not one thing.

(On algae specifically: giant algal cells are a foundational preparation in plant electrophysiology and do carry propagating electrical signals. I have not put a source for long-distance algal signalling into the ledger yet, so there is no entry for it here. That is a gap, not a judgement.)

The other half of a receiver

Everything above is about a transducer — the thing that turns a field into a signal. The peafowl crest is about what sits in front of one, and it is the clearest published case I have found of a biological pre-filter whose contribution can be read off in numbers rather than assumed.

A peacock’s display rattles its train at around 26 Hz. A peahen’s crest — a fan of twenty to thirty feathers, each a tapered shaft with a wide flag at the end — is a mechanical resonator, and its resonant frequency sits inside that band, with filoplumes and their mechanoreceptors at the feather bases ¹⁰.

This is the one source on this page I have read at full text, so it is the one whose findings I will actually state. The crests were measured on a shaker: female crests resonate near 28 Hz, male near 26, with quality factors around 5 and 6.

The number that matters is that quality factor, because at resonance it is roughly the amplitude gain. Being tuned bought a factor of about five. Not a thousand, not a hundred — five. And a Q of five is a broad filter: about a twenty per cent band, which is a wide target for a 26 Hz stimulus to land in.

So the resonator is not rescuing a signal from below the noise. It cannot be: the impacts in question deflect the crest by around nine millimetres, which is visible on ordinary video and sits enormously far above what a mechanoreceptor needs. What the tuning buys is rejection — a narrow-ish window centred on the frequency conspecifics actually display at, which ignores wind and footfalls and broadband racket. The gain is almost incidental.

That is the transferable lesson, and it cuts against the intuition that biology must be doing something extraordinary: a passive biological resonator buys single digits. There is a real ceiling here — wet keratin at body temperature does not make a high-Q cavity. Sharper biological filtering exists, in the cochlea, but it is active: it spends energy, and anything claiming that trick has to say where the power comes from and what noise it adds.

And the honest caveat, which the paper states itself: every one of those measurements was made on dried crests mounted on balsa. No live bird, no nerve recording, no behavioural test. That the crest is mechanically tuned to the display frequency is measured. That peafowl actually use it that way is not, and the authors say so.

Two ways to feel the Earth

Magnetoreception is where the divergence gets sharpest, because the two best-known mechanisms have almost nothing in common.

Some bacteria build chains of magnetic particles inside membrane-bound compartments ¹¹. The cell is then physically torqued into alignment by the Earth’s field. This is a compass needle in the most literal sense: no energy budget, no light, no computation, no nervous system. The physics does the work and the organism goes along with it.

Migratory songbirds are thought to do something entirely different — a light-dependent reaction in a protein in the eye, where the magnetic field influences the chemistry of a short-lived pair of radicals ¹². This is a chemical magnetometer. It needs a photon to start, and it reports the field as a change in how a reaction turns out.

Copy the first and you have built a magnetometer. Copy the second and you have built a spectrometer that happens to be field-sensitive. Same environmental quantity, two instruments with nothing in common — which is a useful thing to have internalised before deciding what “detecting the field” is going to mean for a given design.

What I take from it

Three things, all structural, none of them a finding of mine:

1. A field sensor can be a motion sensor. Two of the three organisms above read electric fields with hairs. Any charged, compliant surface is a candidate transducer.
2. Tuning is worth single digits, and it buys rejection more than gain. The one measured number on this page is a factor of about five. A front end narrows the band the noise arrives through; it does not lift a signal that is already under the floor.
3. Geometry alone can convert a field into motion, with no receptor in the loop at all.
4. Long-distance conduction in biology is not always ionic. Electrons travel too.
5. If an instrument emits, cancelling its own signal is most of the work — and the one animal that does this times the cancellation from the command, not from the recording.

None of this is evidence for anything I am claiming. It is a map of where the good ideas already are, and every one of them was arrived at by something with no access to an amplifier.

Corrections

If any statement above overreaches its source, that is a defect and I want it reported — the whole point of citing by key is that you can go and check. The spider-web paragraph exists because I believed the wrong version myself.

What this post describes is the method — how the questions get picked, built for, and narrowed. It asserts nothing about biology, so it cites nothing. Everything else here is held to a different standard, and the difference is the point.

Build first, narrow later

The order matters, and it is the opposite of how caution usually works.

First, anything goes. An idea is not required to be plausible, funded, peer-reviewed, or even likely in order to be built. Absence of evidence is a reason to construct something, not a reason to decline. Traditional practice, fringe claims, “I want to hear what this sounds like” — all sufficient. The cost of building a thing that turns out to be nothing is a weekend. The cost of never building it is that you cannot tell the difference between a bad idea and an untried one.

Then, narrow. Every constructed thing is aimed at a claim that could kill it, and the conditions that would kill it are written down before the measurement, not after. That machinery is unglamorous and it is most of the work:

• A viability gate runs before anything is built, and asks four questions of the instrument you actually own rather than the one you wish you had. Is the thing you want to read even inside the band the hardware passes? Is the effect above a measured noise floor, in a session a human would sit through? Can the stimulus be produced, and what does it cost? Is there a control that isolates the claim from everything else? Any “no” stops the build. This rule was bought expensively: an entire apparatus was once written, and then the gate was computed, and the arithmetic that killed the build was four lines long.
• Kill conditions are pre-registered. A claim arrives with the list of results that would end it, fixed before data exists. A hypothesis you cannot describe the death of is not being tested.
• Blind reads are structural, not promised. Where a judgement could leak, the analysis is built so the answer is not reachable from it — the truth lives in a file the reading code does not name, and the reveal refuses while any case is still uncalled. “We were careful” is not a control.
• A negative is worth exactly what the test’s power was. A non-detection from an instrument that could never have seen the effect is not evidence of absence; it is a gap in the instrument. Those two get different words and are never merged.
• Refusals live in code. Where a mistake would be invisible in prose, the software raises instead. A predicted value may not be styled as a measured one. A confidence tier may not be argued upward.

What “narrowed” actually means — and what it does not

This is worth being exact about, because the obvious reading is wrong.

Nothing here has been falsified. Not one claim under test has been shown false about the world. What gets established is almost always something about the apparatus — and those are completely different statements.

The ladder in use, roughly weakest to strongest:

The distinction that took a while to get right: a door closed for this bench is not a door closed in science. When a gate fails, the honest sentence is “the amplifier in front of me filters away the thing I wanted to read, and no amount of processing downstream recovers it” — a fact about a $70 board. The underlying question stays exactly as open as it was, and someone with better hardware should go ask it.

Getting this backwards is easy and it is the most common way a research programme lies to itself. An agent working on this corpus once marked several claims falsified on forward arithmetic alone, with no measurement taken. The protocol now forbids it: arithmetic flags, it does not close, and the promotion to falsified needs a measurement and a person.

So the useful output is rarely “this is false.” It is usually:

• here is the number that says this instrument cannot reach it,
• here is precisely what would change that,
• and here is what it would cost.

That is worth publishing. It saves the next person the months, and it is the part nobody writes down.

Full-stack, on purpose

The other half of the approach is a refusal to specialise, and it comes from software. A full-stack developer is not the best in the world at any one layer. They are someone who can follow a problem down — from the interface, through the protocol, into the transport, to the hardware — without handing it off at every boundary, because most real bugs live exactly where the hand-offs happen.

Applied here, one question routinely crosses:

neuroscience · cardiology and electrophysiology · audiology and psychoacoustics · molecular biology and gene regulation · developmental bioelectricity · acoustics and archaeoacoustics · geophysics and space weather · electromagnetics and propagation · signal processing · information theory · statistics and meta-research · materials science · and the history of the practices that got here first

Those are not a list of interests. They are the layers one question passes through. Ask whether a sound can change what a cell transcribes, and you are immediately in acoustics (what leaves the speaker), psychoacoustics (what the ear encodes), systems neuroscience (where it lands), electrophysiology (how it is measured), signal processing (whether the measurement is real), molecular biology (what a durable change would even look like), and statistics (whether you would know). Hand that question off at each boundary and the parts come back individually correct and jointly meaningless.

The cost is real and worth stating: nobody working this way is the strongest person in any one of those rooms. The compensation is that the boundaries get looked at, and the boundaries are where this particular question lives.

What that produces

Open questions, narrowed carefully, with the reasoning and the arithmetic in public — and a clear line between what the world has not told us yet and what this bench cannot currently hear.

The posts that follow are those questions.

Specifications

Checked 2026-08-27.

One correction worth making, because it is easy to repeat. “Dual” refers to the microphones, not the transducers. Every listing I checked says “Built-in Dual Mics”. Two bone transducers, one per side, is the standard arrangement for any stereo bone-conduction headset and is not a distinguishing feature.

The property that decides everything else

A sealed transducer at one ear gives roughly 40–60 dB of interaural attenuation — the signal arriving at the far ear is far quieter than at the near one, which is what makes two ears independent. A transducer on the skull gives close to none. The skull couples both cochleae, so a bone-conducted signal arrives at both at once.

This is not a shortcoming of this product. It is what bone conduction is.

It has a hard consequence for the work I do. My rulebase returns a laterality of undefined — not “left”, not “bilateral-ish” — for any delivery path with under 15 dB of interaural attenuation. A path that cannot deliver a left-ear stimulus must not have downstream predictions made as though it did.

So, for this device:

Two further limits, neither disqualifying. The 20 Hz – 20 kHz figure is an air-conduction-style specification and I have not verified it as delivered through bone on a head — for evoked-potential work the delivered envelope is what matters, and it is unmeasured. And the nearest measured air-versus-bone separation I have is 22 dB, from tuning-fork work on a different transducer for a different purpose.

First impression

Mine, one person, one session, unblinded. Not evidence, and the tier on this post reflects that.

The headphones were extremely easy to get set up, and this is the first time that I’ve used bone conduction headphones, so it truly blew my mind experiencing them. I couldn’t believe how high resolution they were, even though it’s not even sound directly through the ears. Still trying to wrap my head around it. But regardless, this is a great buy.

Also, just as a side note, it’s a really great way to have audio-evoked potentials or something like Brain.fm running during family time so that my ears aren’t covered, and I’m more aware and more present, but can still be running a therapeutic music solution.

That last point is the one I would actually defend, and it is about compliance rather than acoustics: a protocol that can run while you are present with your family is a protocol that gets run. Nothing here measures whether the delivered signal is adequate for the purpose — only that the barrier to wearing it is much lower.

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

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.

Microvolt EEG on a 10-bit converter is only a few counts. It’s quantisation-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

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:

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 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 a broadband RMS figure 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.

What I’ve actually run

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 that as rig validation and nothing else — evidence that board, electrode, amplifier and analysis code together find a real, expected signal. It is 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.

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 published precedent

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 ¹, later extended to combined auditory and visual stimulation ².

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 medial superior olive — the first place in the ascending pathway where the two ears are compared. There’s a far-field potential from phase-locked activity there, the neurophonic.

The constraint comes from the modelling work on that signal ³. I’ve engaged with the released model rather than the paper’s full text, so I’ll state what’s checkable: 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 drive arrived at the nucleus. It does not license cells crossed threshold. Those are different statements and a postsynaptic far-field can’t separate them. Most of the downstream biology I care about 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.

Except for one thing.

The analog ceiling

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

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 fitted to the rig rather than the rig chosen for the question. That’s a reasoning failure, not a hardware problem, and I’d rather name it than have someone else find it.

It’s 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. 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 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 a hardware question.

What I’d like

Not a collaborator, co-author, or funding. A 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?
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. How far does consumer-grade EEG extend before it stops recovering what lab-grade systems get, and where exactly is the boundary? That’s publishable with or without my program attached, and I’d run the tedious conditions.

I have a bench test queued that will probably close a large part of my program — a sustained-response decay measurement. If it fails, most of what I’ve built downstream was wasted effort. 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

It’s a distraction from the bench. The sustained-response test needs a headphone, an electrode, and an afternoon. It’s 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 I can defend is a delivery-verification methods result. What made me start 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 a warm response doesn’t solve my actual problem. Better instrument, corrected montage — neither gets me tissue access and pharmacological control. That’s a different building, probably a different institution, and possibly a different species.

The rule that generates all the others: a reader must be able to check everything here, and an error must not be able to survive a commit.

How this is written

The questions, the claims, and every verdict are the author’s. What gets built, what would count as killing a claim, and what rung a claim sits on are human decisions. The protocol governing this work forbids an AI from promoting a claim or calling one falsified — that requires a measurement and a person.

Drafting and analysis are AI-assisted. Prose, analysis code and measurement pipelines are produced with an AI assistant working inside those protocols, then read and edited by the author. Research assistance runs on a local, zero-retention model on hardware the author controls.

Citations are neither hand-written nor model-generated. Every reference is resolved against Crossref or PubMed by a separate tool and copied in mechanically. A validator rejects any citation typed by hand, any bare DOI, and any key not already in the ledger — so a reference here cannot be invented, by a person or by a model. Where a source is cited without its full text having been read, the article says so rather than describing what it found.

Confidence tiers are copied, not argued. An article may not claim more confidence than the weakest claim it rests on, and that is checked before it can publish.

That disclosure used to live in a collapsed drawer in the footer of every page. It belongs in an article instead, where it can be read rather than dismissed.

The gate

Publishing is blocked by a script rather than by good intentions. It enforces, among other things:

• Cite by key only. A citation in the body is key="levin2018" and nothing else. A bare DOI in prose fails the build. So does an inline Author et al. 2021 — because if a name cannot be typed by hand, it cannot be invented. That rule exists because an audit once found a paper cited for months under an author who does not exist.
• Both directions must agree. Every key declared in a post’s front matter is cited in the body, and every key cited in the body is declared. A source that is listed but never used is dead weight; one used but not listed has no verified provenance.
• Every key resolves in the ledger, which is machine-copied from the research repo. If a key is not there, it does not exist for publication purposes.
• No tier laundering. A post may not assert a confidence tier stronger than the weakest claim it rests on.
• Nothing from a retracted or archived directory, and no self-asserted [VERIFIED] badges — that is a check that passes upstream, not a label anyone applies.

The gate runs again in CI against the committed data, without access to the private research repo. That is deliberate: what this site claims must be checkable from the public repository alone.

What the gate cannot do

It proves the citation plumbing is sound. It cannot prove the prose is honest. A correctly-cited paper can still be misrepresented, and only reading catches that.

So there is a manual pass every time, looking for the specific ways a claim outruns its source: correlation written as causation, animal results written as human results, in-vitro written as in-vivo, a single study written as consensus, an effect reported without an effect size. And the standing rule that an abstract is not a paper — if the full text has not been read, the post may say the work exists but not what it found.

The tools

Everything here is built from open, inspectable pieces, which is the point:

The confidence badges are worth one note. The five tiers are distinguishable by colour, but colour never carries the meaning alone — each badge also ships a glyph and a word, because the palette fails a colour-blindness check past three categories and a reader who cannot see the difference should not be the one who loses information.

How posts stay in sync with the research

A post is not a snapshot that drifts. The chain runs one way:

1. A claim lives in the research repo with its sources, its computed tier, and the conditions that would prove it false.
2. A sync script copies only what published posts actually reference into this repository — the bibliography and the claim tiers, nothing else. It refuses to copy a key that never resolved against a registry, or a claim with no tier, or anything from a retracted directory.
3. The gate then validates the posts against that copied data, and a separate check fails if the copy is stale relative to the research repo.

So a tier displayed on this site is the tier the corpus computed, at the commit the data was copied. If a claim weakens upstream, the post cannot keep the stronger badge past the next sync.

Corrections are amended, never erased. A factual error gets a dated correction appended to the post that made it, and if a claim is retracted upstream it is retracted here in the same change that updates the data. Nothing is quietly edited away.

What this site is not

It is not a results feed. Most of what gets established here is a fact about an instrument — that a given piece of hardware cannot reach a given question — and that is a different statement from the question being answered. The distinction is load-bearing and it gets its own post.

If something here overreaches its source, that is a defect. The whole point of citing by key is that you can go and check.

Note — 2026-08-31

This post replaced a placeholder of the same date that existed to give the publishing machinery something to validate. The original made no claims and cited two sources purely to exercise the citation renderer; both have been removed along with it, and the post is now tiered A as a description of settled process rather than C. It is filed under method rather than audio-evoked potentials, which is what it was always about.