The redesign was done in the open, in conversation with Claude on a design canvas where every version stayed side by side. Then it was built with Claude Code. Six directions, oldest first, and then fourteen passes on the last one:

1. A familiar front door

The first brief was a proper homepage for a research lab, borrowing the feel of product sites like DigitalOcean and Intercom: a big headline, a product shot, and sections for what the lab does. It was built strictly on polarize-ui, the design system these pages already used: the display serif for headings, the mono face for labels, and never bold.

The product shot was a mock recording workbench, with three channels and a time–frequency map. It carried a TASTE badge saying it was an illustration and not a recording. Even as decoration, the page kept the rule every page here follows: a label may never be hidden.

2. Dark

The same page in the design system’s dark palette. It was an easy change, because every colour is a token. It was also the first sign that the lab reads better at night.

Both of the first two versions had a phone layout too, drawn as separate artboards.

3. Brutalism, and specks

Next came a push towards brutalism, after the sound-therapy site Byotōne: fields of animated specks, bracketed mono tags, a huge headline, and a lettered step-through navigation. The specks took several forms: a ring around the headline, a slowly turning cell, a helix, and an evoked-response wave. The page labelled them decoration, not data.

One thing from the reference was deliberately not copied: its all-caps serif. Here uppercase is always the mono face, so the big headings stayed in regular case.

4. No rules

To see what the rules were costing, one version dropped all of them: a condensed grotesque, acid green, 1px grid lines everywhere, and a simulated network of about five thousand cells behind the whole page. Those cells fire on their own, set off their neighbours, and sometimes fire together as waves, synchronised colonies or rhythmic bursts. A live counter in the top bar says how many are firing, and the footer says it is a simulation.

It was the loudest version and the least like the lab, but the cells stayed.

5. The one that shipped

The final version took something from each of the others. It has the dark palette and the design system’s type. From brutalism it took hard rules and square-ish corners, and from the first version it took the charts. The cells moved into a single WebGL field behind the entire page, one that looks like cells polarising rather than confetti. They are small, white, dim, sometimes grouped, occasionally firing together. The stepper and the bracketed tags went, and so did the background fills on cards.

The content changed more than the look. Looking at how other labs present themselves helped, especially Michael Levin’s lab site and his separate essay site, Thoughtforms. They pointed to an order: lead with the work, let the articles report what it found, and give speculation its own quieter room. So the homepage now opens with the process that keeps results honest. Then come the projects, the changelog, the findings, the speculations and the design system.

Fourteen passes on the hybrid

The hybrid itself changed fourteen times before it shipped, one request at a time. These are all of them, in order, each shown where it changed the page.

6. From canvas to components

A design that lives only on a canvas has to be rebuilt by hand for every page, and it drifts. So it went into polarize-ui as components, in both of its forms: a zero-build stylesheet and custom element for plain HTML sites like this one, and React components that render the same classes.

The new landing layer covers the top bar, hero, section heads, process steps, project features, repository cards, the changelog feed, article cards, story cards, archival plates and the footer. The cell field is <ui-cellfield>. It draws nothing a reader needs, and it stops for anyone who has asked their system to reduce motion. The design system’s own checks hold the new stylesheet to the same rules as the rest: no raw colours, never bold, and uppercase only in the mono face.

It shipped as polarize-ui v0.5.6, and this site then adopted it everywhere. Every page now has the new top bar, the new footer and the dark palette, and the speculations page uses the new story cards.

The same components hold up on a phone. The top bar folds into a menu that works without script, the process steps stack with downward arrows, and the story cards scroll sideways.

What was kept

Every version is still on the canvas, and the rules did most of the work. The versions that broke them were useful for finding out what the rules cost, and each time the answer was less than it first looked. These are the ones that survived:

• The type rules. The display serif is for headings, the sans is never bold, and uppercase is always mono.
• The label rule. Decoration says it is decoration. A tier badge can never be hidden.
• Real content only. Every card is built from the site’s own data: posts, essays, repositories and the changelog. A project’s finding is always its own post’s summary, never a line written for the homepage.
• The cells. Only these came from the version with no rules.

The components are in polarize-ui, and the live Storybook shows the whole homepage at desktop and phone width.

scientific-research-scaffold is the result of writing those decisions down once. It is a protocol for how a piece of research is laid out across repositories, and scaffold, a small command-line tool that sets new repositories up in that shape and checks existing ones against it.

What the layout has to protect

Every rule in the protocol exists to keep a result checkable by someone who is not the author, on a machine that is not the author’s. Four things break that quietly:

• A classification baked into a name. A simulator here started life as sim-neural-memory and was later renamed to neural-memory. A prefix fixes a category into a name that outlives it, and renaming breaks every link and path that pointed at the old one.
• Two copies of one thing. A simulator imported as a folder by two studies is two copies waiting to diverge.
• A path on one person’s machine. A result produced through an absolute home-directory path is a result nobody else can regenerate.
• Findings in more than one place. A research corpus mounted inside each study is several working copies of one record, and they drift.

Two kinds of repository

The protocol has two kinds of repository, declared in a manifest rather than in the name:

• A study is one research question and everything used to pursue it: small versioned apps for exploring it by hand, simulators for modelling it, preregistered experiments for testing it, and pinned datasets to test it against.
• A sim is one simulator under adaptive preregistration. Its versions are tags inside one repository, never new repositories.

Neither holds findings. A study’s findings go to the shared corpus, and the study refers to them by claim ID. Everything specific to one organisation (which repository is the corpus, which design system the apps use, where shared tools live) sits in a profile, so the protocol itself stays generic and can be forked.

An app is for exploring and an experiment is for testing. An app answers “is this worth testing?” A number from an app is never a finding; a finding comes from a preregistered experiment.

When a simulator moves out

A simulator starts as a folder inside the study that needs it. It moves to its own repository when one of three things becomes true: a second study uses it, it needs releases on its own schedule, or it is shared outside on its own. Tidiness, size and “it might be reused someday” are not reasons.

scaffold promote-sim does the mechanical part. It splits the folder’s history into its own branch, so the new repository keeps every commit, and the study then pins the new repository by tag like any other dependency.

Strictness that scales, and rules that don’t

How strict a repository has to be depends on what a mistake would cost, and the protocol names four stages. A sketch owes honest labels and a note of what was not done. Once a number is being produced, it owes a preregistration before any run that could be quoted. Once another repository depends on it, or a number from it is quoted, it owes CI, a lockfile and tagged releases. Once someone outside relies on it, it owes a frozen release, a DOI and an external preregistration.

Some things never relax at any stage: the safety of anyone taking part, honest labels, other people’s data licences, and saying plainly what was not done.

What the tool does

scaffold new study renders a template, makes the first commit, installs the preregistration kit, and prints the command that would create the repository on GitHub. It leaves the visibility decision to the person, and it never publishes anything. scaffold new app adds the next versioned app, and scaffold new sim adds a simulator, inside a study or as its own repository.

scaffold check is the part that matters day to day. It confirms that the manifest is valid, that the name carries no prefix, and that the kind in the README matches the manifest. It checks that visibility was decided and dated, that the preregistration kit is installed, and that every external simulator is pinned to a tag rather than a branch. It checks that the corpus is not a submodule, and it warns about absolute home-directory paths. It exits nonzero on any failure, so it can gate CI.

Adopting it without breaking anything

An existing repository already has working code, frozen results and gates that pass, and adopting the protocol must not break any of them. Adoption is additive first: add the manifest and point its entries at wherever things already live, add the kind line and the research register, install the kit, and run the repository’s own gate. It must pass exactly as it did before. Moving folders into their default places is optional and comes later, one commit per move.

Analyses frozen under an earlier preregistration system stay exactly as they were frozen. They are listed in the experiment register with a note saying which system froze them, and the kit is used for new experiments only. Converting a frozen record would change the very thing the freeze exists to protect.

Where it comes from, and what it is not

The rules draw on published practice on organising computational projects, scientific computing, reproducible research, preregistration, adaptive preregistration for model-based research, and research software as a citable output. The protocol lists those sources in its closing section. No published protocol was found for laying out a whole research programme across repositories. The split into studies, simulators and one shared corpus is an extrapolation from that practice, and it should be read as one.

Using it

The protocol is licensed CC BY 4.0 and the code MIT. It needs Python 3.9 or later and git, and nothing else. To use it for your own work, fork it and write a profile naming your own corpus, preregistration kit, design system and workbench; any of them can be left empty. The first release, v0.1.0, is on GitHub at polarizetech/scientific-research-scaffold.

The question

Could a network of spiking neurons, with synapses that tag and consolidate the way published models describe, store several sounds at once and play them back later from its own activity?

The project was set up to fail visibly. Every mechanism has an off switch, every result sits beside the same run with that mechanism removed, and a plain reservoir network is always run alongside as a baseline.

The tape did not work

Sound went in readably: a linear decoder recovered the input envelopes from the network’s activity while it listened. Nothing came back out. At every delay, and with every way of prompting recall that was tried, the decoded output matched the stored sound no better than it matched sounds the network had never heard. The plain reservoir encoded better.

An early result looked like a partial success: which stored sound was dominant at a given moment could be read from the order in which cells fired, 77% of the time where shuffled labels reached 57%. A later code review showed that ordering was measured while the sound was playing and was driven mostly by the input itself. It says the cells followed the input, not that the network kept anything.

The weight changes explain the failure. What the network wrote into its synapses was set by which cells happened to fire hardest, not by what they heard. A weight change predicted from any of twenty sounds the network never heard matched the real one as well as the played sound’s did.

Habituation as memory

The project then asked a smaller question. Repeated sounds evoke smaller responses. Is that fading a memory of the particular sound, or just fatigue?

Four mechanisms were compared in a small auditory network, each with its own switch:

• depletion of the input synapses;
• depression of a synapse only when the receiving cell also fires;
• inhibition that learns, using a rule of the form published in ⁴;
• receptor removal and replacement, as modelled for a single-celled organism ¹.

Two experiments were preregistered: predictions, pass criteria and code were fixed in git before anything ran. Both closed as failures, because some predictions missed. What they showed:

Depletion kept no sound-specific memory. At no delay, from 2 seconds to 90 minutes, did it suppress the trained sound more than a different one. The main reason is in the ear model. At a normal listening level, the least-similar of twenty random test sounds still overlapped the trained one by 0.59 in the simulated auditory nerve, on a scale where 1 is identical. Depleting shared input fibres cannot tell the two apart.

Only plasticity gated by the receiving cell was specific. Depression that needed the receiving cell to fire picked out the trained sound 30 minutes later. The learned inhibition suppressed the trained sound too, and removing that inhibitory pathway undid all of it. Its specificity was limited by the same 0.59 overlap.

A model that drains where the cell does not

The fourth mechanism reproduces a published model of habituation in Stentor, a single-celled organism. In that model, stimulated receptors are pulled inside the cell and then recycled or destroyed, and new receptors are made continuously ¹. In real cells, blocking protein synthesis made habituation faster and memory longer, while untrained cells given the same drug kept responding normally ².

Run with the published model’s own rates, and the paper’s time step read as one minute (the paper states no unit), the block did not do that here:

• untrained responses fell by 32% after 20 minutes and 79% after 90 minutes;
• the extra memory the block produced was at most 1.4 points of suppression.

The published model has basal loss of surface receptors, so a synthesis block drains untrained cells in the model too. There the drain is hidden by a hard threshold: a cell far above its contraction threshold keeps contracting while its receptors run down. A readout without that threshold, like this network’s, exposes it. In this model, the cell’s normal untrained response depends on the threshold, not on the receptor pool. Whether that holds in the organism is a question for an experiment, not this simulation.

Spacing did not help

In roundworms trained at 60-second intervals, long-term habituation followed training split into blocks with hour-long rests. Training all at once lowered next-day responses too, but only the spaced group responded significantly less on the second day than on the first ³.

None of the mechanisms reproduced a spacing benefit. With the same number of presentations, spaced training left less suppression than massed training in the receptor and learned-inhibition models, and no difference for depletion. (The excitatory-depression model also favoured massed training, but a later review found its untrained comparison network drifts during a long schedule, so that comparison is confounded.) Each mechanism has a single decay rate, so earlier blocks simply fade. A spacing effect seems to need a second, slower consolidation stage that none of these models has.

What would change these conclusions

• A test pair with genuinely low overlap in the auditory nerve (below about 0.2), at which depletion does keep a specific trace. The overlap limit is a property of this ear model at this level, not a general law.
• Adding a hard response threshold to the receptor model’s readout and finding the untrained response still drains.
• A single-timescale mechanism that nonetheless favours spaced training.

The project is paused here. Its findings, defects and the next experiments each would need are listed in the repository.

Software makes this cheap. A new variant costs seconds, and the run that finally agrees looks exactly like one that was planned. The fix is the same one experimental science uses: write down what would count as success before looking, and keep a record that shows the order things happened in.

This is the protocol used for model experiments here. It adapts published work on adaptive preregistration for model-based research and on reporting deviations from a preregistered plan. It is written for simulations, analysis pipelines and research software, whoever or whatever is doing the running. The full protocol, with its sources and how deeply each was read, is in the protocol document.

The rules

• A claim, or an exploratory label. Every piece of work, whether an app, a simulation, a tool or a calculator, starts with a falsifiable claim, settled in the author’s own words together with what would count against it. Work the author marks as exploratory may be built without one: it is recorded and labelled as exploratory, nothing it shows counts as a finding, and it gets its claim before its first experiment is preregistered.
• No run before the preregistration. A run whose result a report will cite may not execute until PREREG.md is complete and tagged.
• The environment is part of the plan. The preregistration names the model version it runs against, and a lockfile pins every dependency. Adding a dependency afterwards is a deviation.
• Failure closes the version; it does not edit it. A failed experiment is closed with its failure written up, and the next attempt gets a new experiment ID. Every ID stays in a registry, so a pass after earlier failures is reported as one attempt among several.
• Every departure from the plan is logged, dated, with whether the outcome was known at the time.
• Exploratory work is quarantined. Anything not in the plan lives in exploratory/ and is labelled exploratory everywhere it is mentioned.

What makes it adaptive

Modelling is iterative, and a protocol that forbids iteration gets ignored. The rule is about order, not rigidity.

• A plan may change between stages. Each stage says in advance which result leads to which change. The change is committed and tagged before the run it governs, and logged with the outcome marked as not yet known.
• A plan may also change after a result is seen. It is still logged, with the outcome marked as known, and it can never be used to convert a fail into a pass.
• An experiment that fails stays in the record. Nothing is deleted or cleaned up.

What the preregistration contains

Eleven sections, all written before any scoring run:

• the question, as one mechanistic sentence;
• risky, numeric, directional predictions, each with the result that would count against it;
• pass and fail thresholds for each estimated quantity, with its uncertainty: a pass means the interval clears the threshold, not just the point estimate;
• the outcomes that would close the line of work;
• the model specification, with every parameter tagged as taken from literature, derived, or arbitrary;
• the design: conditions, a fixed list of seeds, controls, exclusion rules, a justified number of runs with a stopping rule, and a sensitivity analysis for every arbitrary parameter. A pass that flips to a fail anywhere in a preregistered range is reported as not robust;
• the environment, down to the execution backend and thread count, because the same seed does not give the same result everywhere;
• what the authors already knew: which target data they had seen, and which targets the model was calibrated to rather than tested against;
• any adaptive stages, every decision the specification left open, and how the plan was timestamped.

The results file answers the plan line by line. It also has to list the patterns the model failed to reproduce, what only worked because it was tuned, and the result of every control.

The record

Git holds the ordering. Each milestone is an annotated tag:

Each experiment gets its own branch and draft pull request. Tagging a milestone posts a receipt to the pull request with the commit and a checksum of the plan and the lockfile. The comment’s timestamp comes from the hosting service, not from the machine that made the commit, so it shows the plan existed by then.

Rules for whoever runs it

The same rules apply to a person at the keyboard, a script, or a coding assistant.

• Before running anything citable, confirm the plan is complete and tagged. If it is not, write it and stop for someone else to review.
• Never edit the plan after it is tagged. Changes go in the deviations table.
• Asked to “make it pass” or “try a few values and keep the best”, do the runs under exploratory/ and say plainly that they are not preregistered evidence.
• State every open decision before the run that depends on it.
• In every report, state how deeply each source was read: full text, abstract, background knowledge, or supplied by the user.
• If anyone asks to skip a step, write the skip into the deviations table first, with the outcome-known field filled in honestly.

How it reaches every project

The protocol is maintained in one public repository, KIT Adaptive Preregistration, and installed into each project with a small command-line tool, kit_ap. It copies the protocol documents and tools into the project’s .agents/ folder and records the exact version in a lockfile. It also writes the rules into AGENTS.md, so coding assistants working in the repository read the same rules a person would.

Some of the rules are backed by code rather than left to good intentions:

• an optional commit guard refuses outputs for an experiment that has no preregistration tag, and refuses any change to a plan after it is tagged (like any git hook, it can be bypassed);
• where the tools support hooks, the current experiment’s plan status is surfaced on every session;
• at the start of a session, the project checks whether the protocol has changed upstream, and reports it without updating on its own.

Limits

This protects the record, not the idea. A well-preregistered experiment on a wrong model is still wrong.

A local git history is not tamper-evident: tags can be moved unless they are signed and protected on the host. The pull-request receipt is a third-party clock but not an archive, since comments can be edited, and it can’t show that nothing ran before the plan. For outside credibility, the preregistration tag should also be deposited with an archive, and the plan should say which.

Nothing stops a new attempt after a failure. The registry only makes the attempts visible, so a reader can weigh a pass against them.

Most of the protocol is rules that people and tools are asked to follow. Only the commit guard and the hooks enforce anything, and the hooks only in tools that run them. The rest relies on the record making a violation visible afterwards.

Update — 2026-09-24

The kit’s repository is now public and linked from the section on how the protocol reaches each project.

After a methods review, the protocol was tightened, and this post was revised to match. It now describes: the registry of attempts, uncertainty in the pass criteria, a justified run count, a sensitivity rule fixed in advance, a section on prior knowledge and calibration (eleven sections, previously ten), and decision rules for adaptive stages. The limits section now says more about what the record can and can’t prove. Previously the post said a pass that depends on an arbitrary parameter does not count; the rule is now that a pass which flips within a preregistered range is reported as not robust.

Update — 2026-09-25

Recording coding-assistant conversations is no longer part of the protocol, and the kit no longer installs it. The post previously said each prompt and reply was posted to the experiment’s pull request and committed beside it; that paragraph and the matching line under “How it reaches every project” were removed. The kit’s repository was also renamed, and its links here point to the new address.

Update — 2026-09-30

The protocol gained a rule that comes before the others: every piece of work starts with a falsifiable claim, stated with what would count against it, before anything is built. Settling it means saying what the claim’s terms mean, how it will be measured, the smallest effect that would matter, and what a test of it assumes. The section above was previously titled “The five rules”. Preregistration now also works per piece of work: a project holding several apps, simulations, tools or calculators gives each its own claim, experiments and versions.

Update — 2026-10-07

The first rule changed. It previously said every piece of work starts with a falsifiable claim before anything is built. Work can now be marked exploratory instead: it is built from a short note of what is being built and what is already known, it is labelled exploratory, and nothing it shows counts as a finding. Its claim is settled before its first experiment is preregistered. Settling a claim before building now means stating the claim and what would count against it. What its terms mean, how it will be measured, the smallest effect that would matter and what a test of it assumes are stated in the first preregistration, not before building.

This is the method used in KIT Retrieval-Augmented Generation:

Pipeline

• Plan: turn the question into sub-questions, search terms and falsification queries.
• Discover: search the available literature providers and record providers that did not answer.
• Acquire: fetch accessible full text through the paper library.
• Index: split the text into passages with stable evidence IDs.
• Retrieve and rerank: combine lexical and semantic search, then select the passages to inspect.
• Extract: record the passage, source, direction, population, study role and a verbatim quote.
• Synthesise: let the model draft only from evidence IDs that were actually retrieved.
• Critique: search again for null results, opposing results and failed replications; do not merely ask the model to reconsider its own draft.
• Verify: check quotes and numbers in code, then judge each claim against its cited passage.
• Render and log: print bibliographic metadata from stored records and save the searches, passages, model versions, verdicts and limits.

Rules

• The model fills bounded, schema-constrained forms. Code controls the sequence.
• A paper’s existence and a passage’s support for a claim are separate checks.
• A quote must occur in the stored passage.
• A number in a claim must occur in its cited evidence, with a matching unit where the field declares units.
• A citation must be an evidence ID from the run; the model does not type the bibliography.
• A passage describing another paper is a pointer, not independent first-hand evidence.
• Verifier disagreement is reported as disputed, not averaged away.
• Failed claims are removed and counted in the limits.
• “No opposing result was retrieved” is a statement about the search, not about the literature.
• Retrieved papers are data, not instructions. Text addressed to an AI reader is excluded when detected.
• A run’s own output is never indexed as a source for a later run.

Extending it

The engine is domain-neutral. A domain package supplies the rules that the engine cannot know:

• which study designs count as evidence;
• which sources are authoritative;
• which units are equivalent;
• which species, preparation or outcome gaps limit generalisation.

The domain is recorded in the run log. Adding a neuroscience, cardiovascular or vestibular extension changes the field rules without forking retrieval, verification or logging.

Limits

This does not make a language model truthful. It reduces the paths by which unsupported claims enter an answer and makes the remaining uncertainty visible.

It does not recover missing figures or supplements, make closed papers available, or guarantee that a verifier is right. When another model writes the answer through the repository’s MCP evidence tools, only the deterministic checks apply (evidence ids, quotes and numbers); whether a passage supports a claim is judged only inside the pipeline. A useful result can still be insufficient evidence.

Update — 2026-09-24

The repository is now public and linked above. Before publishing, the code was brought into line with this description: a claim whose number does not occur in its source is now removed rather than marked disputed, and the MCP evidence tools now apply the quote and number checks too.

The subject is the European green crab (Carcinus maenas), an invasive portunid established on the coast of British Columbia. The question I set out with was narrow and, I thought, tractable: is there any signal a green crab can detect that a native crab cannot — because if there is, you can build it into a trap and catch one species without catching the other.

Twelve routes were screened. Ten are closed. One is real but unreachable with equipment I can obtain. Two survive, and neither of them is a device.

None of this is novel. Every closure came from applying published work or undergraduate physics to a specific question, and in one case the thing I declared missing from the literature turned out to have been published, with public code, four months earlier. The value here, if there is any, is the shape of the negative results and the order they arrived in — so I am writing it down in public rather than leaving it in a private repository. If any of it saves somebody a month, take it.

The constraint that governs everything

A green crab trap that also catches Dungeness crab is worse than no trap. Dungeness supports British Columbia’s most valuable crab fishery, and juvenile Dungeness share estuarine nursery habitat with juvenile green crab. So “attracts green crab” was never sufficient. The requirement was always differential — something the target detects and the bycatch does not.

That single constraint is what killed most of the twelve.

The ledger

Five independent lines of work converged on the bottom two rows. That convergence is the single most useful thing the exercise produced, and it took about eight closures to get there.

Where it started: the electromagnetic branch

The project began with a specific published result: juvenile green crabs exposed to static magnetic fields at strengths relevant to submarine power cables showed a sex-specific response, with females spending more time in exposed zones and males showing no consistent preference.

That is exactly the shape of thing the constraint above demands — a differential response, in the target species, in a channel you can generate with a coil. It is why this project exists at all.

Three things had to be true for it to become a trap, and I could establish none of them.

First, a mechanism. No transducer has been located. The obvious comparison is the spiny lobster, the standard invertebrate magnetoreception system — but the receptor cells 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.

Second, the field map has to reproduce. It does not. No coil geometry I could construct, at any separation or tank bearing, reproduces the published zone table — best fit was 61% RMS error. That is a correspondence-with-the-authors problem, not a finding, and it is recorded as one.

Third, the confound. Helmholtz coils dissipate heat and crabs are thermotactic. Working the published geometry through a thermal model gives, at the strongest field, 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.

One thing argues against the thermal explanation, and it is worth stating because it cuts my way and I nearly missed it. 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 it is the cheapest experiment anywhere in this review.

Verdict: open, and unreachable here. The claim is untouched. What closes it is a replication with a sham coil and a temperature log, which needs a facility and animals I do not have.

And the electric branch is closed, which is a different answer

The original premise was electric, not magnetic — a lure imitating the bioelectric field of buried prey. It is closed on two independent grounds.

The physics. 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 source. Run through a shark’s own measured sensitivity, a natural prey field is readable at a few centimetres. A crab is not a shark.

The biology. No electroreceptor has been identified in any crustacean, and none has been sought in green crab. A behavioural sensitivity figure exists for a freshwater crayfish — a surrogate, from a different order, in fresh water.

That is not a modelling gap. It is the state of the field, and the honest output was to say so rather than to model around it.

And the welfare evidence points the same way. Green crabs have nociceptors in every soft tissue tested, and they learn to avoid a place where they receive electric shock — some dropping a wired leg even though the voltage was set below the single-shock level that causes it. That was contact current, not a field in water, so it sets no threshold. But it means an electric lure for this animal should be expected to repel as well as attract.

Prey signalling: vibration and sound

If the crab cannot be lured electrically, the next question is what its prey actually broadcasts, and whether a trap can imitate it.

The receiver side is better documented than I expected. There is a real transfer function for the mouthpart setae — displacement thresholds, and the coding: spike count scales with displacement amplitude, interspike interval with velocity. That paper is also where I made my first substantive error (below): those setae are tactile receptors for handling prey, not distance detectors, and I spent a while scoring far-field stimuli against them.

The organ that matters for substrate vibration is a joint organ in the walking leg, described in green crab itself, with a measured band. Water motion is a different receptor again; the behavioural response of intact animals to substrate vibration has been tested directly; decapod acoustic detection has been reviewed as a field; an evoked-potential study in the American lobster asks which organ is responsible; and a consolidated review of green crab sensory biology exists.

The transmitter side is where it died. The reference dataset for benthic invertebrate sound gives calibrated source levels for scallops, limpets, urchins and crustaceans. Converting each to particle displacement and solving for the crab’s threshold, not one of the nine sources 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, and the distinction is the whole point of recording it.

Two pieces of arithmetic govern all of this and are worth stating plainly, because they close route 4 without any biology at all:

• Near-field falloff. A small oscillating body in water produces a field falling off as 1/r³ if it oscillates back and forth. The cube law is brutal: a thousand times more source amplitude buys ten times the range.
• 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.

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

Mating: pheromone, and interference

Chemistry out-ranges every mechanical cue by two to three orders of magnitude. Modelled as a plume, prey odour reaches hundreds of metres where the best mechanical cue reaches a couple. 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 electrophysiologically.

The female sex pheromone has been the subject of an identification study, 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 significantly. Noise interfered with completing the behaviour, not with detecting the signal. That dissociation is worth more than either number alone.

So chemistry works, and that is the problem. It is the best-performing channel in the whole review and it is the least selective one: a bait plume is smelled by every crab in the basin. A pheromone lure is better — it targets males of one species — but males are demographically close to irrelevant, and it does nothing about Dungeness bycatch.

Route 5, acoustic reproductive interference, closes on physics rather than on biology, and cleanly. Shallow water is a waveguide with a cutoff frequency below which no mode propagates at all. At the 0.5–5 m depths where these animals actually mate, the cutoff is 159–1594 Hz — above the signal band you would need to mask. You cannot interfere with a signal in water that will not carry your interference.

The selectivity problem, stated properly

By this point the pattern was clear enough to test directly, so I built a screening tool: for a given stimulus, how much arrives at a receptor, in what format, at what range — and a refusal when the number underneath does not exist. That has its own write-up.

Its most useful output was a comparison that cannot be run.

Across eight receptor channels and five species, 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 and both behavioural — a detection threshold for a prey extract and salinity detection measured by antennular flicking. Both are held at abstract only, so those are descriptions of what was measured, not of what was found.

Two filled cells out of thirty-two. And they do not make the comparison runnable, because they are behavioural thresholds in grams per litre while every green crab value is single-unit electrophysiology in micrometres. Putting them in one row would be a category error, not 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.

And where the bands are known, they overlap. Every crab whose mechanoreception has been measured peaks around 100 Hz. There is no frequency at which you can shout at a green crab and not be heard by a Dungeness.

What can be compared is everything except the senses, and there the differences are substantial and well documented: salinity tolerance, depth and zonation, body size, moult timing, tolerance of emersion and hypoxia. Subadult Dungeness use the littoral zone in ways that matter for any comparison drawn from trap data, red rock crab emersion physiology has been measured in situ, settlement and emigration behaviour has been studied in green crab, and predatory capability has been compared directly between green crab and red rock crab on a native clam.

That list is the beginning of routes 11 and 12. Selectivity does not have to come from the signal. It can come from the calendar.

Predators and parasites

The complete answer to “why not biological control” is one paper. The host specificity of Sacculina carcini, the parasitic barnacle proposed as a green crab agent, has been tested against native Pacific crabs, and the result is why nobody deploys it. I am not going to characterise the numbers here — I hold it at abstract level — but the conclusion is not controversial and it is the standing reason route 9 is closed.

Route 10 has the only landscape-scale evidence in this entire review, and it is not usable. A recovering sea otter population shows a strong negative space-and-time relationship with green crab, with crabs persisting highest where tidal restriction excluded the otters. That is the best result anywhere in the predator literature — and the southern sea otter has been absent from the Salish Sea since 1915. It is an argument for removing tidal barriers, not a lever anyone can pull.

So I built a register of everything else that might eat a green crab in British Columbia: 24 agents, each carrying a status so the empty rows stay visible. The tally is the finding — 1 demonstrated, 1 failed, 3 with a measured diet record, 4 plausible-untested, and 15 with no data at all. Five of the top six candidates are present in the Salish Sea. None of the five has demonstrated evidence there.

Four survive three filters at once — present where the crabs are, capable of taking one, and attached to a step someone could take this year: giant Pacific octopus, marine-foraging river otter, Pacific staghorn sculpin, and the sunflower sea star.

Building out their biology reordered them, and the reordering is the point. Ranked on being present and cheaply samplable, the river otter comes first. Ranked on how strong the dietary prior actually is, the octopus does. Those are different questions and collapsing them into one ranking is how a project talks itself into the wrong survey. The sunflower sea star has the strongest structural case on the list — a recovering predator with funded recovery programmes already running for their own reasons — and the weakest dietary case, because its entire measured feeding ecology is urchins and molluscs. Both statements are true at once.

What I can state, because these are my own search results rather than somebody’s findings: no green-crab record exists in the diet literature for any octopus species; there is no green-crab diet fraction for river otter anywhere; there is nothing published on Pacific Northwest fish predation on green crab; and I located no crustacean predation record at all for the sunflower sea star. The diet numbers behind the ranking sit in the working notes at abstract level, where they belong.

And none of the four is measured on the size class that matters. Which brings up the thing that kept recurring.

The two that survive

Route 11 — gear. Trap entrance modification has the largest demonstrated effect of anything in this exercise. It is a cable tie and a sinker. Nothing the sensory workstream produced comes close.

There is also a peer-reviewed integrated population model for this species with public code, fitted on a multi-year removal time series — which I confidently declared did not exist, and which puts shrimp traps well above the trap type used throughout British Columbia on per-trap efficiency. I would not quote a single multiplier from it: the estimate shifts noticeably depending on which auxiliary dataset the model is fitted with, which the authors report themselves. The defensible statement is that a paired deployment — the two trap types fished side by side, same days, same sites, same soak — would pin the number down cheaply with existing gear.

Route 12 — place and schedule. Selectivity from the calendar rather than from a device. Two axes came out strong: trap the structured habitat rather than the open flat, and fish the daytime flood tide, because green crab move up-shore on the flood regardless of light while the Dungeness incursion is nocturnal. Others came out weaker, and the axis most often cited — tidal height alone — was downgraded by a paper showing subadult Dungeness use the littoral zone too.

It costs nothing, because every axis is a decision about where and when to set gear already in the water. It is falsifiable now, at a few hundred trap-sets, with the kill condition written in advance.

And here is the case against it, stated as strongly as I can. It is bycatch reduction, not control. The demographic work says only heavy size-selected removal of juveniles drives the population down; composition does not. And the gear cannot execute that lever. The smallest crab in one dataset of 17,615 animals was 21.1 mm across the carapace. Settling juveniles are 1.5 mm. The decisive size class and the available gear are mismatched, and no amount of siting fixes it.

That mismatch is also true of every predator in the register. The size class that decides the outcome is unmeasured for gear and for predators alike. If one sentence from this whole exercise is worth carrying forward, it is that one.

Things that kept being true

• Physics beats biology at range, every time. Six of the twelve closures came from arithmetic done before any literature search — near-field falloff, waveguide cutoff, displacement-versus-pressure, ohmic heating. Doing the arithmetic first would have saved weeks, and the one branch where I built the apparatus before computing the gate is the branch that wasted the most time.
• A cue that works is usually a cue everyone can hear. Selectivity and efficacy pulled against each other in every sensory route. The channel with the best range was the least selective; the most selective idea was the one that would not propagate.
• Absence of a measurement is not absence of the thing. Repeatedly the honest output was unavailable rather than absent — a prey-sound corpus filtered before the relevant band, native species never tested, a receptor never sought. Collapsing those into a null manufactures a confident negative out of a test nobody ran.
• The literature on the invader is far better than the literature on the natives. Green crab is a laboratory animal. The commercially important Pacific crabs it competes with are not, and every cross-species comparison in this review foundered on that asymmetry rather than on modelling.
• Removal suppresses locally and does not extirpate. That is the consistent finding across jurisdictions, including a documented case of stage-specific overcompensation — the hydra effect — following adult-only removal, whose possibility had been assessed in advance. Early detection has a much better record.
• The temperature literature is where the actual predictive signal is. Adult thermal tolerance does not discriminate sites — it is broad enough to make the whole coast suitable. The discriminator is larval recruitment temperature, with warmer winters adding a second brood and a modelling result finding a large increase in larval arrival for a small temperature rise. The adaptive-genomics literature for this species is likewise about temperature, not about reproduction or sensory biology — an observation about what the literature is about, which is a claim I can make from titles and abstracts.

The method, and the errors it caught

The part of this I would actually defend is not any of the twelve verdicts. It is that every constant is pinned by a test that fails if the figure is misquoted, every number carries a tag saying whether it was measured, derived, assumed or unavailable, and a source read only at abstract may be cited as existing but may not have findings attributed to it.

That last rule is unpopular with my own drafts. It caught five violations in the first version of the companion post, and it is why several paragraphs above name what a study measured rather than what it found.

Here is what those rules caught. I think this table is the most useful thing on the page.

Two patterns in that table are worth more than the individual entries.

The first: most of these errors ran in the direction that made the result look better. Not one of them was random. That is what a bias looks like from the inside, and it is the argument for tests that pin numbers rather than vigilance.

The second: a test can hold a wrong number in place. The “zero native channels” claim survived two days after I knew it was wrong, because a test was asserting it, with a comment beside it saying the value was known to be stale. That is worse than an unchecked number — it looks checked. And when I finally corrected it, the fix nearly introduced a new error: the function answering which channels are comparable across species tested only whether a cell was marked “measured”, so filling the two Dungeness cells would have certified a category error as a comparison. A check that counts statuses will happily certify nonsense.

What would change the picture

• A thermocouple in a Helmholtz coil tank, during a replication with a sham coil. Settles route 2 either way, and it is the cheapest decisive experiment in the review.
• A measured audiogram for green crab, Dungeness and red rock crab on one rig. It does not exist. It is the single measurement that would change the most, and if the bands turned out to differ substantially, most of the mechanical conclusions above would need redoing.
• A paired deployment of two trap types, same days and sites. Settles the gear question with equipment already in the water.
• Any diet study at all on the four shortlisted predators — two of which need no collection permit, because middens and scat are lying on the ground.
• Gear that reaches sub-20 mm crabs. Nothing in this review addresses the size class that the demography says decides the outcome. That is the open problem, and I do not have an answer to it.

If any of this is useful to somebody working on a real protocol, it is yours — that is why it is here rather than in a private folder. The working notes, including the numbers I have held back above because I have only read their abstracts, are more detailed than this post and I am happy to share them.

What runs where

The model and the voice synthesis run outside containers because containers on Apple Silicon cannot reach the GPU. Everything else is in containers and bound to loopback.

Checking the claim

Zero retention is stated as something to verify, not something to trust. Three checks:

1. List live connections during a conversation. Nothing in the chat path should connect anywhere but the local machine.
2. Turn off Wi-Fi. Chat, retrieval, speech-to-text and text-to-speech keep working. Only phone access and journal pushes need a network.
3. Read the switches. Open WebUI’s telemetry, analytics, update check and web search are off in the environment file. Qdrant’s usage statistics are off in the compose file.

The Qdrant switch was missing until 2026-09-15. Qdrant ships with usage statistics on, and the stack never turned them off, so one of its two containers was sending anonymised usage data while the documentation said telemetry was off. Found while checking this write-up against the code rather than the README. The switch is now set, and the repository’s self-check fails if it is ever removed.

Where it deliberately is not zero-retention

Four exceptions, each off unless switched on, each documented in the file that makes the request.

None of these sees the conversation. A search query is still data, and those services log requests like any website.

A tier, not a slogan

“Zero data retention” describes where the stack is pointed, not the code. Every model, embedding, speech and voice backend is a URL. The deployment declares which tier it is in:

Only local can be verified from the machine itself. It is the default.

The journal

The assistant can make three kinds of write — a finding, a theory update, a session log — into a separate git repository.

• Only on an explicit request (“log that”, “save this”). Otherwise the write is refused and the assistant says nothing was written.
• One commit per write, append-only. No amend, no rebase, no force-push, no whole-file rewrite. The history of a theory file is the history of the thinking.
• The container never runs git. The tool writes the file and queues a request; a committer on the host turns each request into one commit. A crashed container cannot leave a repository half-written.

Literature: the model only asks

In the literature tool the model’s whole job is to turn a request into search terms and pick a function. The archive ranks the results. The tool draws titles, authors, abstracts and retraction notices straight from the archive’s records into a panel. The model is told the reader can already see the panel and not to restate it.

A citation on screen never passed through the model, so the model cannot misquote it. Tools that place a language model between the reader and the paper are deliberately not used. The archives are a fixed allowlist of four hosts, enforced in code.

Memory is the design constraint

On a 16 GB machine the model shares memory with the containers, the embedding model and the operating system. If the machine is swapping, the model is too big, and no quantisation setting recovers it. A smaller model that fits in memory beats a larger one that pages.

Two further rules from running it:

• Avoid models that reason before calling a tool. Journal writes and searches are tool calls; a model that thinks at length before each one turns a voice exchange into a wait.
• Raise the context length. The runtime’s default context is small enough that a system prompt plus a few tool definitions can silently push the tools out.

Why a firewall ships with it

A laptop can hold a publicly routable address with no router in front of it. Then every service listening on all interfaces is on the open internet — including a model server that must listen beyond loopback so the containers can reach it.

On the machine this was built on, that was not hypothetical. Counted from the model server’s own access log: 1,009 successful inference requests from 160 addresses outside any private or tailnet range, 5,985 outside requests in all, between 15 August and 9 September 2026. The last is on 9 September, the day a default-deny firewall for the public interface was committed. Local requests continue in the same log after it.

So the firewall travels with the stack: loopback, the tailnet and the container bridge stay open, the public interface does not. A network check reports in one line whether a newly joined network is hiding the machine or exposing it.

Limits

• An 8B-class model on a laptop is not a frontier model. The trade is capability for being usable with material that should not be uploaded.
• macOS on Apple Silicon only, 16 GB minimum. The GPU paths assume it.
• Tailscale’s coordination service sees device metadata, not content.
• The four exceptions above are real egress whenever they are switched on.

The parallel, and whose it is

A swarm choosing between two nest sites and a cortical population choosing which way dots move have been given one formal description.

That paper dates the idea to “at least Hofstadter (1979)” in its own introduction, and describes itself as a first step toward a common framework — one built on how a system’s components interact rather than on what they are.

The experimental work usually paired with it concerns inhibitory signalling between scouts. Its title and framing put the comparison to brains in front, rather than leaving it to a reader.

What the description covers

Three properties, per the 2009 treatment:

• Populations, not individuals. Interacting populations accumulate noisy evidence for competing alternatives.
• A threshold on one population. The alternative whose population crosses first is the one taken.
• A threshold that moves. Raising it buys accuracy and costs speed. A parameter in the cortical model; a headcount at a site in the ants.

The claim is stronger than resemblance. Under particular settings the cortical model reduces to a diffusion process implementing the sequential probability ratio test — for a two-way choice, the procedure minimising decision time at any given error rate. The analysis asks which colony models do the same, and treats the answer as a prediction about how colonies should behave if deciding optimally.

Past two options the analysis changes.

This is also where the literature meets the best-of-N problem in distributed and robotic systems — the same problem stated without the bees.

Where its own authors say it stops

The colony must go and get its evidence. In the cortical task, evidence for both alternatives arrives at a fixed equal rate. Scouts discover sites stochastically, leaving an explore-or-exploit problem the cortical version does not have: assess what is known, or look for better. The authors call it a hybrid of a bandit problem and a minimum-decision-time problem, and say nobody had formalised it.

Consensus is not foraging. The paper states that colony emigration is optimal consensus decision-making and “differs from the problem of distributed resource intake maximization that colonies tackle during foraging.” Quorum figures and the cross-inhibition result come from nest-site selection during swarming. Foraging recruitment runs on different machinery. A description that moves between the two is describing two systems.

Two extensions, neither established

Recorded as framings, not findings.

A modulatory layer. The models have recruitment and inhibition, and nothing setting the regime those run in. Bees have candidates.

Whether these belong in the architecture as a separate axis — gain rather than content — is untested. The third cuts across the tidy version: if a targeted inhibitory signal moves a diffuse modulator, then selecting among candidates and setting the regime are not cleanly separable.

A write step. Queen pheromone has been examined against expression of genes responsible for epigenetic modifications in worker brains.

Placing that as the durable-change step is an analogy across levels. The level mismatch is real: the bee side is role-level — dancer, follower, scout — and the neural side is cell-type-level.

Two observations that cut the other way

The signal is shaped by its audience. Removing bees from the dance floor, and separately replacing adults with bees too young to follow, was used to test whether dance information content depends on who is watching. Directional and distance precision tracked follower numbers; mean waggle-run duration and distance per run are reported unchanged.

A variance effect, not a bias effect. A population whose output precision depends on its readership is not a one-way sender.

No offline replay on the bee side. The dance is a waggle run, then a return loop alternating left and right. The loop repositions; no reversed waggle run is reported. Compressed reactivation of place-cell sequences in both directions is a large neural literature. The bee-side absence is absence of reported evidence, not a study that looked.

What would make this wrong

• Follower recruitment determined almost entirely by instantaneous dance properties, with no contribution from prior dance-floor state.
• Simultaneous dances that do not measurably interact through competition for a finite pool.
• Cross-inhibition in swarms and cortex producing qualitatively different dynamics under matched conditions.
• Thresholds turning out to be observer cutoffs rather than nonlinear population transitions.

One caution on the neural side. Reaching for cortical disinhibition as the mechanism of attention and then borrowing it imports an unsettled claim.

On sources

Two of the papers above were read beyond their abstracts: the 2009 framework, in the sections carrying its conclusions, and the audience study, its results section. The rest are entry points. Where this post says a study examined or tested something, that wording is load-bearing — it marks pointing at a result rather than having checked it.

Not a simulator. A simulator implies you turn a crank and get behaviour out. This is narrower: 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 refuses when the number underneath does not exist. The most common output is a refusal.

The design rule

Every constant carries its provenance: measured in this species, measured in a surrogate, derived from physics, or assumed. That sounds like bookkeeping. It did the work.

The tags caught six errors already in the model, every one from reading a summary instead of a source. Two examples:

• A trap statistic entered as “entry success rises from 16% to 59%.” Those are two different measurements — one the fraction of attempts that succeed, the other a percentage increase in catch from a modification. The second does not continue the first, and the best modification was not the one being quoted.

• A receptor threshold was carried as a single value where the study reports a range. The single number was the low end.

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

Mechanical senses — the measured part

The strongest ground, because there is a real transfer function for this species. Mouthpart setae recordings give displacement thresholds in micrometres and, more usefully, the coding: spike count scales with displacement amplitude, interspike interval with velocity, and roughly half the cells are directionally sensitive.

That detail is what made the first version wrong. Those setae are tactile receptors for handling prey, not distance detectors, and far-field stimuli were being scored against them. The tool now raises a scope flag on any far-field verdict from that row.

The receptor that is for water motion is a different one — sensory hairs on the chelae, studied in crayfish. A joint organ in the walking leg has been described in green crab itself, the response of intact animals to substrate vibration tested directly, decapod acoustic detection reviewed as a field, an evoked-potential study in American lobster asks which organ is responsible, and a consolidated review of green crab sensory biology exists.

Where the abstract-only rule bites: of the seven sources in this section, two were read at full text. For the other five: they exist, and the question they asked is stated, not what they found. So the surrogate threshold the tool uses for the water-motion channel is a number I cannot attribute, and it is tagged as a surrogate for that reason.

The physics that decides most of it

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

Displacement is not pressure. Particle displacement equals velocity over 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.

Together these explain why the acoustic route keeps failing and the chemical route keeps winning: a plume falls off as 1/r, and the olfactory threshold is low.

What prey actually emit

The reference dataset for benthic invertebrate sound gives calibrated source levels for scallops, limpets, urchins and crustaceans.

Converted to particle displacement and solved against the crab’s threshold, none 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 prey detection, and the crab’s band was removed before analysis. A null there is unavailable, not absent.

The promising cue is not a sound. Porewater pressure signals from infaunal activity have been recorded directly in sediment, and the irrigation timing of three tellinid bivalves — including a Pacific Northwest clam that is real green crab prey — has been measured.

A slow hydraulic signal, not an acoustic one, and the only prey cue here 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. Two anatomically separate systems do different jobs — olfaction on the antennules, contact chemoreception on the dactyls, where the gustatory organs of this species have been characterised electrophysiologically.

Abstract only, so the threshold separation between the two — which is what the tool runs on — is a sourced parameter in the model and is not stated here.

One deliberate refusal: 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 has been identified, and later work with it produced the 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 — while the time taken to respond to the pheromone did not change significantly. 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

The two channels refused outright, and the refusals are its highest-confidence outputs.

Electric. No electroreceptor has been identified in any crustacean, and none sought in this species. Not a modelling gap; the state of the field. The physics is unambiguous: prey bioelectric fields measured at microvolt scale with the electrode under a millimetre from tissue, and the authors state the voltage from invertebrates was too weak to record away from the source.

A behavioural sensitivity figure exists for a freshwater crayfish — a surrogate, different order, fresh water.

The tool quotes named surrogates and will not produce a crab number.

Welfare, which is a different question from detection. This species has nociceptors. Acetic acid on every soft tissue tested — eyes, antennae, antennules, claws, leg joints — produced a central nervous response, longer-lasting and lower in amplitude than a touch.

And green crabs learn to avoid a place where they are shocked. Given two shelters, with a 10 V, 180 Hz, 200 ms shock delivered every 5 s through wires on the legs in one of them, they learned to walk left or right to reach the other. The voltage was set below the level at which a single shock made a crab drop a leg, and some dropped a wired leg anyway.

Both read in full. That shock was contact current through the legs, not a field in water, so it gives the tool no field threshold and changes no number above. What it changes is the expectation: current is aversive to this animal, so an electric lure should be expected to repel as well as attract. The welfare limit the tool enforces — crayfish forced electrotaxis, about 4 V/m — marks where current forces movement, not where it becomes unpleasant.

Magnetic. One directly relevant result: juvenile green crabs exposed to static magnetic fields at strengths relevant to submarine power cables, females spending substantially more time in exposed zones while males showed no consistent preference.

A real, sex-specific response in this species — and a static-field spatial preference, not a compass, and not a receptor. Nobody has located the transducer.

The obvious confound, stated plainly. 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 no behavioural thermal-discrimination threshold for this species could be found to compare against. The model returns a number and refuses a verdict.

One weak argument against the thermal explanation: 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. Confound and 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, the standard invertebrate magnetoreception system.

Here the abstract-only rule bites hardest: all three read at abstract, so their results are not characterised. What can be said is about the shape of the literature rather than its contents — the receptor cells 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

The most informative part, because the comparison cannot be run.

Across eight receptor channels and five species, green crab has measured values in five channels and a described-but-unmeasured organ in two. For Dungeness, red rock, graceful rock and yellow shore crab: two measured thresholds, both in Dungeness — a behavioural detection threshold for a prey extract, and behavioural salinity detection measured by antennular flicking.

Both at abstract only, so those are descriptions of what was measured, not of what was found — the model carries them with the same tag. (A third record sits off the matrix entirely: dactyl chemo- and mechanoreceptor recordings in the kelp crab, not among the five.)

Two filled cells out of thirty-two, and the comparison still cannot be run — for a 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. Not the same quantity, so 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 documented: salinity tolerance, depth and zonation, body size, moult timing, tolerance of emersion and hypoxia.

The first was read at full text and the rest at abstract, which is why this section names questions rather than answers.

The genetics, and what is actually there

Looking for a genetic basis for the reproductive advantage — some identified difference explaining 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 plasticity.

Read at abstract only, so their findings are not characterised. But that the surveyed literature concerns thermal physiology and not reproduction or sensory biology is an observation about the literature, and it can be made.

So the honest account of why green crab out-reproduce the natives is life-history and tolerance, not a gene: 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 clutch. Note 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 in one place — body size — because no measured threshold in this species differs by sex, and the tool will not invent one.

The value is not in the numbers. It is that they cannot drift: every constant is pinned by a test that fails if the figure is misquoted, and in one session those tests caught four errors that had entered from secondary summaries of paywalled sources — which is how this literature is normally read.

The most instructive failure was mine, not the literature’s. The claim 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 stale. That is worse than an unchecked number, because it looks checked.

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 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: a measured audiogram for Carcinus maenas, Dungeness and red rock crab on one rig, showing receptor bands substantially different from the surrogate values used here. Most 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 erasing anything above.

• The native-comparison section originally said “two measured chemosensory thresholds in a single species and nothing else.” The two are now cited by key, one 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.
• “What the tool is for” 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.

2026-09-16. One addition, nothing erased: the electric section now carries the green-crab welfare evidence — nociceptive responses and learned shock avoidance — with a note that neither supplies a field threshold, so no number in the tool changed.

Starting observation: an ECG is not only a record of a heart contracting. Identity, sex, age and health have all been read out of one. If that holds for a single signal, the question is what every biological signal carries, and whether the same structural features carry it each time.

The signals catalogued

Six, named by source tissue rather than application. Representative textbook ranges, not hard bounds.

The useful part was unexpected. Laid out together, the six separate on properties that have nothing to do with physiology: how periodic, how sharp the events, where the spectral weight sits, how much DC. You can tell them apart without knowing the tissue.

That is either an interesting fact about biological signals or an artefact of how they are all recorded. Which one is not known here.

The information axes

What kind of information is a biological wave known to carry. Each row is an entry point to a literature.

Identity. Attributing a trace to a specific person is a mature field for ECG, with a parallel line on individuating signatures in EEG.

Sex and age. Both have been read from short clinical recordings.

The discrepancy between an age estimated from the signal and a person’s actual age has itself been studied as a marker.

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

Intention. Muscle activity decodes motor intent; cortical activity carries potentials preceding movement rather than following it.

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

None of those papers was read at full text. They are cited as entry points; what they exist to study is stated, not what any of them found.

The reframing

Building the list changed what the project looked like.

It had been treated as signal processing — extracting a feature from a trace. Laid out together the axes look like a taxonomy problem instead.

Identity, sex and age barely change. State and intention change minute to minute. Health drifts. Those are not the same kind of quantity, and calling them all “information in the signal” hides the most important distinction between them.

That came from making a table, not from running anything.

Waves with nothing alive in them

A reader that finds rich structure in biological signals has to be asked whether it finds rich structure in everything. So the same reader gets pointed at waves with no organism in them:

• Tide gauges — two stations, 61 days. A blind read recovered the principal lunar semidiurnal period and separated the moon’s contribution from the weather’s.
• Solar flux against cosmic-ray flux — 14.6 years. Recovered the known anticorrelation, but only given the full cycle. That makes it a result about how much data you need.
• Seismic, infrasound, ocean sound, geomagnetic field — recovered the secondary microseism and the daily solar-quiet variation, blind.

The standing rule on that log:

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. The structure being found might be a property of the reader rather than of life. Two of those runs did embarrass the instrument — one exposed silent bugs at low sample rates, another showed a standard statistical null invalid on deterministic signals.

Sound, light, water

Mapped, not tested.

The structural properties in question — how loud, how bright, how periodic, how consistent the repeating unit, whether a slow rhythm modulates a fast one — 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 because a tide is well enough characterised that the reader had nowhere to hide.

A correspondence you can write down is not one you have shown. Applying these characteristics across modalities is, until tested the way the cardiac work was, a hypothesis carried by analogy.

What would make this wrong

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