Coherence Is Not Evidence
A supply chain method for LLM-assisted research — documented through the Glitterex mystery.
This essay documents a methodological failure common in LLM-assisted research: the substitution of narrative coherence for evidence. Using the Glitterex supply chain mystery as a case study, it shows how widely circulated but incorrect explanations emerge from discourse convergence — and how a corrective method based on physical supply chain tracing produces a structurally derived, falsifiable answer. The broader implication: effective use of LLMs requires intervention at the level of question formation, not output evaluation.
In 2018, a New York Times journalist visited Glitterex, one of the world's largest glitter manufacturers, at their factory in Cranford, New Jersey. She asked a simple question: who is your biggest client? The manager's answer was instant. "No, I absolutely know that I can't tell you." When pressed for a reason: "Because they don't want anyone to know that it's glitter."
The internet has been trying to answer that question ever since. Across Reddit threads, podcasts, TikToks, and journalism, a rotating set of explanations has emerged: boat paint, military applications, currency security, consumer products.
Every one of these answers is wrong — and exactly the kind of answer a large language model will produce, because it learned from the same discourse that produced them.
This essay documents how to arrive at a better answer: not by refining the narrative, but by abandoning narrative entirely and following the physical supply chain.
The Trap
Large language models are trained on human-generated text. Human-generated text tends toward narrative coherence. The boat paint theory is coherent. The military theory is coherent. The currency theory is coherent. Each fits some subset of the available evidence. Each has been written about, linked to, discussed on Reddit, and therefore lives in the training data as a plausible answer.
Coherence is not evidence. But to a model trained on text, coherence is the closest available proxy for truth. The most discussed answer becomes the most plausible answer. Six years of internet discourse about boat paint creates a signal that looks, from inside the training data, like consensus.
"Coherence is not evidence. An LLM will hand you the internet's favorite wrong answer dressed up as research."
This is the trap. Not that LLMs lie. The trap is that they optimize for the answer that fits the existing discourse — and if the existing discourse is wrong, they will reproduce that wrongness fluently and with apparent confidence. The only way out is to stop asking what the internet thinks the answer is, and start asking what the physical system actually does.
Six Years of Confident Wrong Answers
The Glitterex question has three documented constraints from the original reporting. The biggest client operates at massive volume. The client does not want the association with glitter known. And the product shows something — Lauren Dyer's exact words were "you'd be able to see something" — but you would not recognize it as glitter.
Every popular theory fails at least one of these constraints when examined carefully.
| Theory | Why it fails the constraints |
|---|---|
| Boat / marine paint | Fails recognition constraint. Metallic boat paint is publicly known to contain reflective particles. "You would never guess it" — people already know boats sparkle. |
| Military explosives | Fails volume constraint. Taggants use trace quantities, not bulk inputs. Glitterex's biggest client is their largest buyer by volume. |
| Currency / passport security | Fails category. Holographic security films are precision-engineered foils, not particulate flake. Different material, different manufacturing chain entirely. |
| Toothpaste | Fails timeline. Microbeads were banned from personal care products in 2015 — three years before the NYT interview. |
| Luxury sand | Fails secrecy constraint. There is no reputational risk in sand. |
Notice that applying the original constraints — all three, simultaneously — eliminates every popular theory. This is not sophisticated analysis. It is the basic discipline of holding the actual evidence in mind while evaluating proposed answers. An LLM that has absorbed thousands of documents about boat paint will not spontaneously apply this discipline. That requires a human in the room who refuses to let the narrative settle.
The First Correction
The investigation began by producing exactly the output described above: a coherent narrative built on internet consensus. The currency and hologram theory emerged early. The origin story — Henry Ruschmann cutting mica for the Manhattan Project, glitter as a byproduct of nuclear weapons research — created a satisfying arc from classified origins to classified present use. It was wrong. The researcher identified the category error: holographic security film is not glitter. It is a precision-engineered product manufactured to exact diffraction specifications. The physics overlap does not make them the same material.
Stop identifying end products. The question is not who benefits from the secret. The question is: what does the physical movement of this material actually look like?
This is the first methodological lesson. The researcher's role in an LLM-assisted investigation is not to evaluate the model's conclusions. It is to evaluate the model's question. When the question is wrong — when it asks what the internet thinks instead of what the system does — no amount of additional research will produce a better answer.
The Wall That Told Us Something
The next methodological move was correct in principle and instructive in practice. The researcher directed the investigation toward shipping manifests — publicly available customs data that records what is being shipped from where to whom. If Glitterex's biggest client is receiving bulk shipments, that paper trail exists.
The manifest approach hit a wall. The public manifest databases — ImportGenius, Panjiva, ImportYeti — aggregate ocean freight data. Glitterex is a domestic manufacturer in Cranford, New Jersey. The data exists — but inside private freight systems, not public databases.
The wall itself was informative. If the biggest client's shipments are not appearing in international freight databases, the relationship is probably domestic. A domestic buyer receiving bulk glitter from New Jersey by truck is not a government mint or a military contractor. It is an industrial manufacturer with a domestic supply chain.
This is the second methodological lesson. A dead end in research is not a failure. It is data. The shape of what you cannot find tells you something about what you are looking for. The manifest wall told us to stop thinking internationally and start thinking about the domestic industrial supply chain.
Supply Chain Logic
The investigation was asking the wrong question. The question had been: which end product contains glitter? The correct question was: who buys glitter as an input material, and what do they do with it before it reaches an end product?
These are not the same question. The first assumes Glitterex sells to whoever makes the thing you eventually hold in your hand. The second recognizes that most industrial materials pass through multiple transformation layers before they reach a consumer product — and at each layer they may be renamed, reformulated, and embedded into something that bears no visible relationship to the original input.
Glitterex's own product documentation made the transformation layer explicit. Their materials are marketed for use in injection molding, extrusion, and masterbatch — the industrial processes by which additives are incorporated into thermoplastics. The Alu*Flake product line is specifically described as ideal for masterbatch extrusion. This is not a consumer-facing application. This is a raw material input for the companies that manufacture plastic compounds.
The investigation had been looking at the bottom of this chain and working upward. The correct approach was to identify where in the chain the material transformation happens — where glitter stops being glitter — and recognize that the buyer is at that transformation layer, not at the end of it.
The Rename
Once the transformation layer was identified, the secrecy mechanism became obvious. The masterbatch and color concentrate industry does not buy glitter. It buys metallic effect additives, pearlescent colorants, special-effect masterbatch, chrome-alternative materials, and effect pigments. The material is the same. The name is different. By the time the material has been embedded in a thermoplastic compound and distributed to an injection molder, it is a resin with specific optical properties — not a craft supply.
Companies operating in the masterbatch and color concentrate sector publicly describe products that perform this exact transformation. Ampacet markets chrome-alternative masterbatch creating "sparkle effects" or "brushed metal effect" in molded plastic. Avient markets metallic-effect and special-effect masterbatches for molded parts. RTP Company markets special-effect compounds for thermoplastics including metallic and pearlescent effects. Teknor Apex explicitly lists "glitter" among special effects in its plastics portfolio — confirming that the industry recognizes glitter as a functional input material, not merely a craft supply.
This is why Lauren Dyer said "you'd be able to see something." The optical effect is present in the end product. You can see the shimmer, the depth, the metallic quality. What you cannot see — what your brain does not register — is that the material producing that effect entered the supply chain as aluminum-coated PET flake from a factory in New Jersey.
And this is why "they don't want anyone to know that it's glitter." Not because of stigma, not because of national security, not because of any of the narratives the internet generated. Because the specific flake formulation is a proprietary competitive advantage. Ampacet does not want Avient knowing which exact Glitterex specification produces their particular metallic effect. Avient does not want RTP knowing their formulation. The secrecy is competitive, structural, and entirely mundane — and it operates at the transformation layer, not at the brand level.
The Answer
The primary buyers of Glitterex material are masterbatch manufacturers, plastic compounders, and color concentrate suppliers — not end-product brands. These companies purchase flake as a raw input, reclassify it as an effect additive, embed it into thermoplastic compounds, and distribute at massive scale. This conclusion is derived from supply chain analysis and publicly available product documentation. It remains unconfirmed by Glitterex or any named company. It is falsifiable.
The buyer class fits all three constraints simultaneously. Massive volume — masterbatch is a bulk industrial commodity. Secrecy — formulation is a competitive trade secret within the color house industry. Visible but unrecognizable — the optical effect is present in the end product; the origin material is not.
The "secret" is not a hidden company. It is a supply chain layer where the material gets renamed before it reaches anyone who would associate it with glitter.
The Lesson
This investigation took several hours. For most of those hours, two large language models — Claude (Anthropic) and ChatGPT (OpenAI) — were generating plausible, coherent, well-sourced wrong answers. The currency theory. The hologram theory. The boat paint theory. Each had internal logic. Each fit some subset of the available evidence. Each was wrong.
The investigation produced a correct answer for one reason: a human researcher who refused to let either model settle into a narrative. Who pushed back when the question was wrong. Who redirected the investigation from end-product speculation to supply chain architecture. The models contributed real value — searching databases, synthesizing product documentation, mapping the chain. But the methodological discipline came from outside the models. It had to. They were too well-trained on six years of boat paint discourse to find the exit on their own.
"The researcher's role in an LLM-assisted investigation is not to evaluate the model's conclusions. It is to evaluate the model's question."
Applied pressure in a research conversation does not mean skepticism of every output. It means targeted intervention at the point where the model has slipped from following the evidence to generating the most coherent available narrative — and redirecting to the physical system instead.
The internet had already done the research. Six years of it.
The failure was not informational.
It was methodological.
The Glitterex mystery was not solved by better research. It was solved by better questioning.
This essay documents a research session conducted on April 7, 2026. The conclusion — that masterbatch manufacturers and plastic compounders are the most likely Glitterex buyer class — is a structural observation derived from publicly available product documentation and supply chain analysis. It has not been confirmed by Glitterex or any named company. The methodology is replicable. The conclusion is falsifiable.
