Somewhere in your bibliography there may be a paper that was never written.

It looks perfect. Real-sounding authors — maybe authors who actually exist, who actually work in that field. A plausible title. A journal you've read. A DOI with the right shape, the right prefix, the confident string of digits. Every formatting convention honored. It is, in every visible respect, a citation.

It just doesn't point at anything. Paste the DOI into your browser and you land nowhere. Search the title and you find echoes, near-misses, nothing. The paper is a ghost, a statistically probable citation, which is exactly what a language model is built to produce and exactly what a bibliography must never contain.

If you write with an LLM anywhere in the loop: drafting, summarizing, "just cleaning up the related work section”, you have probably shipped one of these, or come close. This isn't a hypothetical failure mode anymore. ArXiv now bans authors for a year when submissions contain unambiguous AI-generated errors, hallucinated references named explicitly. Recent sweeps of major ML venues keep finding invented citations in accepted papers — work that made it through review with a ghost in the references. The reviewers who catch them do it the artisanal way: paste the DOI, squint at Google Scholar, repeat.

That moment — a reviewer discovering your ghost before you did — is the cheapest professional embarrassment there is to prevent. So I built the thing that prevents it.

citestamp.com — free, in the browser, no account.

Who builds a citation checker?

I'm an independent researcher. No academic support, no lab, no library subscription that unlocks the grown-up citation indexes. I came to research sideways: years in healthcare, then teaching myself to write software, then discovering that the questions I cared about: information theory, complexity, how systems integrate, didn't require anyone's permission to work on. Just time, secondhand hardware, and an unreasonable tolerance for doing things the long way.

Being unaffiliated teaches you something about scholarly infrastructure most academics never see: nearly all of it assumes you have an institution behind you. The citation databases that could tell you whether a reference is real sit behind subscriptions priced for universities. The tools that watch for retractions charge monthly. Even seeing your own paper's place in the citation network can cost more per year than I spent on the computer I'm typing this on.

Meanwhile the actual public record (Crossref, DataCite, OpenAlex) is right there. Open. Queryable. Free. The raw material for checking whether a citation resolves to a real work has been public infrastructure for years. What was missing was a tool that used it honestly and gave the result away.

So: CiteStamp. It launched this week. Here's what it does and, just as important, what it refuses to do.

Three checks, in your browser, nothing uploaded

Paste a bibliography — BibTeX, a LaTeX thebibliography block, or a plain reference list — into citestamp.com/check, and every reference is checked three ways:

1. Does it resolve? Each reference is looked up against Crossref, DataCite, and OpenAlex — the public registries of record. A reference that resolves to nothing in any of them has the signature shape of an AI-invented citation. That's the ghost detector.

2. Does it agree with its own DOI? The subtler failure: a real DOI attached to the wrong paper. The DOI resolves fine — to something other than what your citation claims. A human skimming the bibliography will never catch this. A registry lookup catches it instantly.

3. Has the publisher marked it retracted? Retraction, withdrawal, expression of concern — publisher events, stated as publisher events. The famous demonstration: run a real published pediatrics paper from 2007 with 277 references through CiteStamp, and it flags exactly one — Wakefield 1998, marked retracted by The Lancet. A citation that was clean when you drafted your introduction can rot before your defense. CiteStamp also keeps a free watchlist that re-checks your papers daily for new publisher notices, so you hear about it before someone else tells you.

A real 2007 paper's reference-health page: 277 references, one flag — Wakefield 1998, marked retracted by the publisher.

That page in the screenshot is itself a feature nobody else ships: any published work can carry a living CiteStamped badge. A dated, embeddable summary of whether its references resolve and none has been retracted, re-checked daily. Not a one-time seal; a pulse. Put it in a README, a CV, a journal page.

And here is the architectural decision I care most about: your manuscript never leaves your browser. The parsing happens client-side. The only thing that goes over the network is a public identifier — a DOI, a title query — sent straight from your browser to the public registries. No account. No upload. No server holding a copy of your unpublished draft. You don't have to take my word for it: open your network tab and watch. Almost every other tool in this space wants your document on their server. Some retain it. One prominent one has no privacy policy at all. I think asking researchers to upload unpublished manuscripts to a stranger's server to check for trust problems is a joke that writes itself.

The same engine lives wherever you write: a Chrome extension (Manifest V3, no remote code), a Zotero plugin, a Word add-in, and continuous checking as you type in Overleaf.

And if you sit on the other side of the desk (à la reviewing manuscripts, editing a journal) the same check runs in reverse: paste a submission's reference list and screen the whole thing in about ten seconds, in your browser, nothing stored. The ghost hunt reviewers currently do one DOI at a time becomes a paste.

What CiteStamp will never tell you

CiteStamp will never call a paper "true." It will never call one "fake." You won't find the words verified or fact-checked anywhere in the product, and that's not timidity — it's the entire design philosophy.

Tools in this category increasingly sell verdicts: AI-powered confidence scores, "95% accuracy," a green checkmark that means... something. A confidence score is a black box wearing a lab coat. CiteStamp's checks are deterministic: a reference either resolves in the public registries or it doesn't; it either matches its DOI or it doesn't; the publisher either posted a retraction notice or didn't. Every result is a receipt you can independently re-derive, not a judgment you're asked to trust.

That's what the name means. "CiteStamped" is a state, not a verdict: every reference in your bibliography has been checked against the record and holds up, and it stays that way because the watchlist keeps checking. It doesn't mean your paper is right. It means your citations are real, consistent, and standing; which is the part a machine can honestly attest to.

The open graph — and the part where I sign my name

Checking references is defense. The other half of CiteStamp is offense: an open citation graph you can see, query, and write on.

Drop any DOI into citestamp.com/graph and watch that paper's citation neighborhood assemble itself — a live force-directed map of what cites it, what it cites, what connects to what. The whole inferred graph, roughly fourteen million edges, is CC0. Public domain. No login, no institution, no meter. The commercial citation indexes charge universities five figures for a view like this and lock independent researchers out entirely. Literature-mapping tools charge monthly subscriptions to draw pictures over open data, and none of them can tell you whether a citation is real, resolving, or retracted while they do it. Ours draws the picture and runs the checks, free, because it's built on the same public record it shows you.

And the record is built to outlive its operator: the signed log is append-only, mirrored publicly, and independently replayable. If CiteStamp vanished tomorrow, the graph wouldn't.

My own paper's neighborhood — signed claims in gold, machine-inferred edges in steel.

And you can walk it. Click any node and that paper's own neighborhood fans out — then another, then another, until you're three hops deep in a corner of the literature you didn't know existed, watching the map grow as you go. It's discovery by relationship instead of keyword roulette: not "what papers contain these words" but what cites this, what it built on — and, as signed claims accumulate, what supports it and what refutes it. It's genuinely hypnotic. I built the thing and I still lose time wandering it.

The graph has two strictly separated layers. Machine-inferred edges, mined from the open registries, render in steel. Signed claims render in gold: a human put their ORCID on a typed assertion that one paper cites, supports, refutes, or extends another, onto a public, append-only, CC0 log. That's post-publication peer review with a spine — not a comment thread, not karma, but attributable claims on the permanent record.

One rule is enforced everywhere and it's non-negotiable: papers, not people. You can sign that a result fails to replicate. You cannot touch the person who published it. Every claim targets a work. The record accumulates argument, never pile-ons.

You can also claim your own papers, sign your ORCID onto your own work, and put your stamp on it. No institution's permission, no waiting to be indexed by someone's proprietary database. For unaffiliated researchers, this is the part I built most selfishly: you put yourself on the map.

For the AI on your side of the desk

The same failure mode runs in reverse: if you're building with LLMs, agents, RAG pipelines, research assistants — your system is one plausible-sounding paragraph away from citing a ghost too. CiteStamp exposes its graph and its checks through an API and an MCP server, so your agent can check whether a citation resolves before it writes it. The MCP endpoint is keyless, point a client at mcp.citestamp.com/mcp and go; no signup, no token, rate-limited but open. The antidote to AI slop turns out to be something AI can use.

For the people who run the pipes

One more audience I built this for: the small end of publishing. Preprint servers, independent journals, and small presses can screen every submission's reference list through the same public API; the big publishers run integrity pipelines built for the big publishers, but this one is open to everyone. If you run a venue and want invented references caught at the submission gate instead of in public, the pipe is already there and the price of entry is zero.

The honest economics

Everything I've described above is free, without caps: the checker, the watchlist, the graph, claiming and signing with your ORCID. Not free-as-teaser. Free because a citation check that only well-funded labs can afford doesn't fix the problem. The ghosts don't discriminate by institution.

There is one paid thing. Pro costs $35 a year and detects nothing that free doesn't. It automates: continuous checking as you type in Overleaf instead of pasting, auto-enrolling everything you cite by DOI onto the retraction watchlist instead of adding papers by hand. If you'd rather paste and click, the free tier is the entire detection engine, forever. (For comparison, tools in this space run $72 to $240 a year, for less checking and more uploading.)

And the numbers, since I promised disclosure: CiteStamp launched this week with zero users and zero paying customers. It's built solo, it runs lean on Cloudflare's secure edge, and it doesn't need to hit a growth target to keep existing. I'm not selling you traction. I'm showing you architecture and asking you to try it.

Try to break it

Here's the ask, and it's small:

1. Paste a bibliography into citestamp.com/check — especially one an LLM helped write. See what comes back.

2. Drop your DOI into the graph and look at your paper's universe. If you post the screenshot, tag it, I want to see it.

3. Tell me what breaks. It's week one and it's one person; there will be rough edges, and I'd rather hear about them than not. I'm at info@citestamp.com.

The second signature could be yours. Get your paper CiteStamped before your reviewer does it for you.

~ If you know someone who may enjoy reading this article, please share ~

Subscribe now

Support open, independent research

Take a photograph, a screen of TV static, and the single word "TRUE." Match them on density. Now run all three through the one number — "organization," "integration," pick your branding — that keeps getting floated as a meter for whether something is conscious.

Same score. All three. They even keep the identical 4/13 of their connectivity. One is a human face. One is noise. One is a fact about the world. The number can't tell them apart.

That's not a calibration problem you fix with a better dataset. It's structural. A substrate-neutral scalar is content-blind by construction. It was never going to see the difference, because the difference isn't in the magnitude, it's in the shape.

So I stopped arguing about it and went looking for where the information the number throws away actually lives.

Three substrates, one test

I pre-registered the whole thing. Wrote down what would count as a win and what would count as a loss before I touched the data. No moving goalposts.

A language model. Whether a statement is true or false is sitting right there in the residual stream — a plain linear probe pulls it out at 83% AUC. The famous scalar? Blind across all 29 layers. The truth was in the structure the number discards.

An anaesthetized brain. You can decode which person a recording came from out of the connectivity alone. The scalar sits at chance. Identity lives in the structure, not the number.

Sleep. There's a recurrence signal that survives even after you statistically subtract the scalar out; every fold, across five stage contrasts. It refuses to collapse into one number.

Now the part I could've buried and didn't. My own favorite idea going in was that recursion (a system folding back on itself) would make the structure measurably sharper. On one substrate that came back marginal. On the next, flat null. It's in the paper, reported exactly that way. A result you have to talk yourself into isn't a result, and I'm not going to dress it up to make my hypothesis look better.

So here's my challenge

Anyone who still thinks a mind reduces to one number: take your scalar. IIT’s Φ, PCI, whatever you trust. Residualize it out of your data, fold by fold, and see if anything that matters survives without it. If it does, your number was never sufficient. It was riding on structure it couldn't see. The test is in the paper. The code is public (https://github.com/existencethreshold/recursive-existence-threshold). Go.

The conclusion fits in a sentence: whatever "conscious enough to count" turns out to mean, it cannot be a single global number. What the number leaves out isn't the substrate. It's relational structure. The multiply-realizable kind a transformer's residual stream and the brain's thalamocortical loops both happen to carry.

I didn't come to this from a lab. I came to it from healthcare, watching consciousness arrive and leave at bedsides for years before I had any math to hang on it. No university, no grant, no permission.

The paper's free and open. Read it, break it, tell me where I'm wrong: https://doi.org/10.5281/zenodo.20566382

If this is the first you're hearing any of it, the four papers it grew out of are collected in the paperback and ebook Foundations of the Existence Threshold. That's the, well … foundation of it all.

Then tell me in the comments: if meaning had to ride on one number, which would you pick? And what would it miss? I read everything. Restack it if it made you think.

~ If you know someone who may enjoy reading this article, please share ~

Subscribe now

Support open, independent research

The book is out today. Foundations of the Existence Threshold — four papers, in order, in one volume.

Here's why that matters, and it isn't the reason you'd guess.

The four papers have been public for a while, and you can read every one of them free. What you couldn't do until today is read them as the single argument they actually are. The connective tissue, the part that shows you the throughline, didn't exist until I wrote it for this book.

So let me show you the throughline. It's the whole reason to read the thing.

Paper one asks when a pattern persists, and answers with one number across eighty cellular automata, no misses. Paper two asks what that number loses when you force a pattern across a dimension, and gets the same answer every time: 86%. Paper three proves why it's 86%. It falls out of counting neighbors on a lattice, 4/13, exact, no error bars, and then it holds on the attention maps of real language models. Paper four takes the number off the grid completely and runs it through three decades of markets, a thousand geomagnetic storms, and fifty sleeping brains. Zero tuning between them.

Each paper stands on its own. Line them up and a result appears that none of them states alone.

The same number breaks in two completely different ways.

Force a pattern across a dimension and the number is destroyed. 86% gone, and it doesn't care what the pattern means. A true statement and a confident hallucination lose the exact same amount. Geometry is blind.

Push a living system to its edge instead, a market into a crash or a brain into deep sleep, and the number isn't destroyed. It's redistributed. Integration collapses, differentiation surges, the total holds. The pieces swap and the books still balance. And now it does care what the pattern means: a market and a brain write the same total in opposite directions.

One process destroys. One redistributes. Same number, two fates, and you only see both when the four papers sit on the same table. That's the book.

That's also the honest pitch. The papers are free if you just want to check a single result. The book is the whole record in one place: four papers, three bridge essays that connect them, the data, the figures, and the failures, laid out so a stranger can follow the argument end to end. And it's a real, citable object on a shelf, which is something independent research almost never gets to be.

If you want it:

• Read it. Foundations of the Existence Threshold — paperback · Kindle

• Verify it. Every paper, with its full review trail, is free at icsacinstitute.org.

• Subscribe for the working notes, not just the finished record.

  Subscribe now

  
  

Erik Hoel’s Dumbo Could Already Fly shows OpenAI’s Erdős result was already sitting inside public ChatGPT 5.5 if you knew how to ask. The new model didn’t produce new capability. It removed friction. Hoel names the moment: the post-benchmark era. Capability claims nobody’s bothering to replicate, dressed as scientific reports.

His diagnosis opens a question worth taking a step further. We have no measure of what an architecture can do that’s independent of how skillfully it got asked. Benchmarks score outputs. They can’t score the substrate. That’s where Dumbo’s feather actually lives; within the absence of any instrument that tells the elephant from the trick.

Anthropic’s “Mythos” cybersecurity disclosure has the same shape as the OpenAI Erdős claim: output behavior under one prompting regime, sold as new latent capability. Neither is falsifiable without independent replication, and the field’s stopped doing that work seriously. “Capability improved” means “we got better at asking, in a setting we didn’t show you.” The cure may already exist. Mechanistic interpretability is the only program asking what circuits can architecturally do, independent of how they got elicited on the day. It’s still treated as a boutique safety subspecialty. It’s the only honest capability metric on the table. Flip that, and “Mythos” becomes falsifiable. Leave it, and every safety claim rests on the same sand as every capability claim. They’re the same problem.

The move underneath is substrate-independence. A real capability measure has to describe the system as an information-processing object: what it integrates, what it differentiates, what it dissipates, without caring what any particular output says. If the measure depends on the model’s answer, you’re measuring elicitation. If it depends on the circuit, you’re measuring capacity. The line is sharp. Nobody’s forcing the field to draw it.

This is the question I work on. The Dynamic Existence Threshold (DET) framework measures substrate capacity as an integration-differentiation balance, and the math doesn’t care what the substrate is. It’s certainly not the only answer. It’s just one that takes the demand seriously. Mech interp does this work inside neural networks. DET does it across substrates. Hoel just named the problem. The coalition’s been in the room.

Until measuring architectures is foundational instead of boutique, capability papers and safety papers are nothing but hand-waving and stage magic. A result the audience can’t tell from the apparatus. Falsifiability collapse isn’t a benchmark problem. It’s a substrate problem we keep solving with prompts.

~ If you know someone who may enjoy reading this article, please share ~

Subscribe now

Help support open, independent research

** Cover image: AI generated **

A family stands in the doorway of a room. The face staring back at them from the bed is the face they have known for decades. The person who used to live behind that face is not there anymore. The body keeps eating, keeps breathing, sometimes keeps talking. The person is gone, or going, or comes back for ten minutes on a random Tuesday and then is gone again. Anyone who has sat with Alzheimer's in a family member knows the specific weight of that room. It is not grief in the normal shape. It is grief while the person is still in front of you.

I worked several years in long-term care as a nursing assistant. Not the best pay, but truly meaningful, and I loved it so much I pursued further education and training to become a vocational nurse. Later, nursing home administration. Eventually, ICU admissions and administration. Same question, different rooms.

The ICU is the acute version. Bodies that keep running on machines while the person stopped being there hours or days ago. Then the other thing, the one nobody on shift gets used to. Terminal lucidity. A patient who has been somewhere else for weeks comes back sharp and present, talks to their son or daughter, says the thing they needed to say, and dies that night. Whatever was missing for those weeks was, briefly, back. And then it wasn't.

After enough years across that arc, the question stops being mystical and starts being structural. What part of a person IS the person? Where does it live, physically? When does it stop being there? Why does someone with Alzheimer's stay recognizably themselves for a long stretch and then, on some unmarked day, stop? Why does one person wake up from a coma the same person, and another person wake up as someone else? It is not "where did they go." It is "what configuration has to hold for them to be here at all."

If I'm honest the question was there before any of this. Long before. Weird things happened when I was a kid that I couldn't fit into any standard categories. The home I grew up in offered us many strange occurrences. What would be referred to as the "paranormal." I dislike that word, paranormal; and I’ve always wondered if what happened there had a scientific explanation. That maybe what we encountered was not only real, but somewhat normal. Something complex, yet mathematical. Something measurable, yet dynamical. The work in long-term care and the ICU just gave the question a vocabulary and a clock.

The question kept reformulating until it went math-shaped. Complex systems. Information theory. Integration. Differentiation. The conditions under which a pattern stays itself across time. I taught myself the engineering. Four papers in, and the framework has a name. The Dynamic Existence Threshold. The mechanism may be general. Cellular automata. Neural networks. EEG data. Transformer activations. Solar storms. Markets right before they come apart. Same form, every time. Some patterns hold their shape. Others don't. There is a clear, measurable line.

The work has an institutional home. I founded the Institute for Complexity Science and Advanced Computing, ICSAC, to publish open and to serve as the citable anchor for cross-substrate work like this. Papers live on Zenodo, CC-BY, free to read and free to break; and I didn't make ICSAC just for me — just because I've been rejected by more journals than I can count — I made it for the vast number of those like me. Independent researchers. Citizen scientists. Those whose work is unaffiliated with any institution or big agency, but is just as rigorous. Those who may be lacking a specific niche word, or term, but understand and demonstrate proficiency in their research. Most importantly, for those who simply can't afford APCs and Author fees. The code for the ICSAC's AI-assisted peer review program is always available for public inspection on GitHub and HuggingFace. I saw a pattern with journal publishing that should no longer be ignored, and ICSAC is my attempt at a fix.

The research happens in the margins of running 3Rivers WebTech, the small tech consultancy I operate here in Fort Wayne, Indiana, while attempting to dodge the eye-rolls and sassiness of a teenaged daughter.

This newsletter is for the work that isn't ready to be a paper. Or, isn't professional enough to be a research paper. Dispatches from somewhere in the middle. Where math, computing, small business ownership, dad-life, and the kid asking why the moon influences ocean tides, all sit at the same dinner table. Consciousness, complexity science, AI, fatherly advice, and the math of what persists. That last clause is the spine.

The reason any of this matters: the people we love come apart, one way or another. Slow in a long-term care room. Fast in an ICU bay. Eventually, completely. Understanding the shape of that, what the configuration is, when it holds and when it doesn't, might be the only honest thing we can do about it. The rest is just being there in the doorway.

Support open, independent research

If this resonated, forward it to one person who’d want to read it