Hook

I was auditing my own feed archive — 1,140 items collected over ninety days across eleven crypto publications — when a headline stopped me. It sat between a note on stablecoin reserve rules and a rollup upgrade post, and it read like this: an esports team had won a closed qualifier, two maps to nil. The item carried a category tag for a tactical shooter. The tournament series named in that headline is, by every public record I could find, a MOBA property. Two classification errors inside one sentence, on a site whose entire commercial premise is that it knows the difference between these things.

Nothing about the result itself is interesting. It is a sports score, and scores are the least durable form of news. What is interesting is that the item passed through a syndication wire, a content management system, a tag taxonomy, an RSS feed, two aggregators, and at least one newsletter before it landed in front of me. Not one of those layers flagged it. A single wrong field propagated through six systems without friction, and every system downstream treated the wrong field as authoritative because the system upstream had emitted it.

That is the thing worth auditing. Not the score. The pipe.

Context

Bull markets do something specific to publishing economics. Advertising inventory expands faster than genuine reporting can fill it, and the gap gets closed with volume. Volume is cheap: wire copy, press release rewrites, syndicated feeds, algorithmic aggregation. The marginal cost of one more item in a feed is close to zero. The marginal revenue of one more indexed page is not. That asymmetry does not require anyone to be cynical. It only requires the incentive to be visible on a dashboard, and it always is.

I spent the 2017 cycle watching this happen in real time. Back then the content farms were human. I remember arguing with people who were assembling technical-sounding posts out of whitepaper fragments and calling it analysis, and I remember thinking the problem was opinion. Too much of it, too loud, too unearned. So I went the other direction. I spent three months auditing the foundational code of the Ethereum Classic fork, specifically the immutable ledger mechanisms, and I filed twelve technical critiques on GitHub. Nine of them were about bugs. Three of them were about the governance philosophy embedded in the hard fork decision — the part nobody wanted to write about, because governance is not a bug you can patch.

Those twelve critiques were my education in a specific lesson: the moment everyone agrees something is a fact, nobody checks it. The hard fork was a fact. The immutability principle was a fact. Both were contested, and both got flattened into slogans within a month, and the flattening happened inside channels that were nominally technical.

That was 2017. The problem then was too much opinion and too little code. The problem now is too much text and too little provenance. Same failure, different layer.

Here is what makes the current cycle structurally worse. In 2017, a human being had to sit down and write the bad post. Today, the volume problem has been automated on the supply side, and the aggregation problem has been automated on the demand side, and the two automations are now optimized against each other. A feed that pulls from a wire is not exercising editorial judgment. It is executing a rule. And rules do not have judgment to exercise — that is their virtue, and it is also the hole in the floor.

So when I say I found an esports score filed under crypto categories with a wrong genre tag, I am not describing a mistake. I am describing the output of a system operating exactly as designed. The design has no step where anyone asks whether the item belongs.

Let me be precise about what I actually pulled apart, because the details matter more than the anecdote.

In my archive, that item had no canonical URL pointing to an original publication. It had a downstream mirror. It had no author entity — not a missing byline, but a missing structured-data node, which means no aggregator or summarizer downstream could resolve an author even if it wanted to. It had a category field inherited from the feed, not assigned by an editor. It had a publish timestamp with no modification timestamp, which means there is no record of whether anything about the item was ever corrected. And it had no disclosure of its origin — no line saying "this came from a partner feed," which is the single cheapest transparency measure available and the one most consistently omitted.

Five failures. Any one of them alone would be survivable. Together they mean the item is unfalsifiable in practice. You cannot trace it, you cannot attribute it, you cannot date its correction, and you cannot tell whether a human ever read it before publication.

Silence is the loudest audit. In an audit, the most informative thing is not what the system logs. It is what the system has no field for.

Core

Classification is a trust claim, and trust claims compound

Every taxonomy is a protocol. This is not a metaphor. A category tag is a claim about the world, emitted by one party, consumed by others, and acted upon without re-verification. That is exactly the structure of an oracle. And like an oracle, its failure mode is silent: the consumer has no way to distinguish a tag that was reasoned from a tag that was inherited from a broken upstream field.

Now follow the propagation. My feed pulls the item. My screeners bucket it. A summarization layer reads the bucket and writes a paragraph. A newsletter quotes the paragraph. And somewhere, a model training corpus absorbs all of it, because caches do not distinguish between verified and unverified text — they distinguish between crawled and not crawled.

A mis-tagged item does not stay mis-tagged. It becomes a fact in every system that never had the ability to check it.

This is the oracle problem with the serial numbers filed off, and the crypto industry should recognize it instantly, because we solved the analogous problem in a different domain and then forgot the lesson applied elsewhere. We spent a decade building smart contracts that are deterministic and verifiable, and then we bolted on price feeds from third parties, and then we discovered that the contract does not care whether the oracle is right. It executes. The contract is honest. The feed is not. The composition of the two is a lie with perfect settlement guarantees.

Content is the same architecture. The classification layer is the oracle. The distribution layer is the execution environment. And the reader is the contract that settles against whatever input it receives.

There is a second-order effect that is worse than the first. Once a taxonomy is known to be unreliable, everything filtered through it becomes unreliable in the same direction — not randomly, but systematically. If a publication tags by genre incorrectly, then any analysis built on "let's look at what this publication covers under this tag" is measuring the tagger, not the subject. I have made this mistake myself. I once built a small dashboard tracking Layer2 coverage volume as a proxy for developer activity, and it took me three weeks to notice I was actually tracking which editors liked the word "rollup." The dashboard was accurate. The inference was garbage.

The instrument had no error bar, and I treated the absence of an error bar as an error bar of zero.

That is the trap. Not bad data. Data of unknown provenance presented with the same confidence as data of known provenance.

Why cryptographic provenance does not automatically fix this

My instinct, given what I do, is to reach for signatures. I want to be honest about where that instinct fails, because I have watched it fail at scale.

Between 2021 and 2023, provenance got financialized. The idea was simple and correct: attach a durable identifier to a creative work, record its origin, let the identifier travel with the work. What actually shipped was a speculative asset class wearing provenance as a costume. The identification mechanism became the product. People were not signing their work to establish authorship; they were signing it because the signature was tradeable. And once the signature was tradeable, the incentive shifted from "prove I made this" to "prove this specific instance is rare," which are entirely different goals that happen to share a hash function.

When provenance becomes an asset, provenance stops being a record and starts being a market. Markets optimize for the record's price, not its accuracy.

I am not dismissing the technology. I am dismissing the assumption that attaching a cryptographic primitive to a problem will solve the problem, when the problem was never about the primitive. The primitive was fine. The incentive around the primitive was the failure.

So let me describe what I actually built in the years after, and what I learned by watching a small group of artists and writers use it, because the design lessons were not the ones I expected.

We built a standard for signed authorship — a way for a creator to attach a verifiable claim of human origination to a piece of work, with the signature anchored periodically and the content addressed by hash. Five developers. Open source. The premise was straightforward: as generated content approaches the cost of zero, the ability to distinguish human origination becomes a scarce good, and scarcity goods get priced.

The first thing we learned is that creators wanted the signature but did not want the ledger. Almost uniformly. The moment a creator understood that an on-chain record was permanent and public, a significant fraction of them asked how to make it unlinkable to their real identity. This is not vanity. This is safety. A journalist in a jurisdiction with press restrictions cannot have a permanently linkable authorship graph, and a system that requires one will simply be unused by exactly the people who need it most.

The second thing we learned is that revocation is not an edge case — it is the feature. People sign things and then change their minds, or get threatened, or discover that the context in which they signed has shifted. A provenance system without revocation is a surveillance system with better branding.

The third thing, and the one that matters most for this article, is that cryptography can prove who signed. It cannot prove who is human. Those are different claims, and only one of them is cryptographic.

When we say "proof of human intent," what we are actually proving is control of a key, at a time, by an entity with a history. The humanity part is not derived from the math. It is attested by cost — the cost of building a history, the cost of maintaining an identity, the cost of being wrong in public repeatedly. Humanity is a social claim that we are back-filling with cryptographic evidence. That is an honest description of the system, and it is much weaker than the marketing, and I would rather say so than let the marketing stand.

Which brings the whole thing back around to the misfiled esports brief. If a signature only proves key control, then a signed misfiled brief is still misfiled. The signature verifies the pipe, not the content. And the pipe, as I established, is not where the error entered.

Why you cannot simply put provenance on-chain and call it solved

The reflexive answer is to anchor everything. Hashes, timestamps, authorship claims, correction history. Put it on-chain, make it permanent, make it cheap.

Permanence and cheap are the two words that do not fit in the same sentence. This is where the current infrastructure conversation becomes relevant, and where I think a lot of people are building against an assumption that will not hold.

Post-Dencun, blobspace is cheap. That cheapness is not a constant — it is a snapshot of a demand curve at a particular moment. Blobs are designed to be pruned. Data availability is not data permanence. Loading a provenance layer onto blobspace means your authorship record has a half-life measured in weeks unless something else carries it, which means you have not actually built permanence — you have built a very fast, very cheap way to lose a record permanently.

And blobspace will not stay cheap. The fee market logic that governs every other resource in this system governs this one too. Demand for data availability scales with the number of rollups and the number of users they serve, and both of those curves are steep, and neither of them is hypothetical. Any system that prices its long-term commitments against today's blob cost is making a two-year bet on a market that is structurally designed to repriced upward. Rollups that assumed the current fee level as a floor for their unit economics will discover, on a timeline they did not choose, that their cost basis moved. Not because anyone did anything wrong. Because the resource is finite and the demand is not.

The correct design is boring. Anchors, not storage. Merkle roots committed periodically, the actual content held off-chain in content-addressed storage, verification happening at the edge by readers who care. That design has been available since roughly 2016. It has not been adopted at scale because it has no business model, and I will get to that.

The subsidy parallel that nobody in crypto wants to hear

Here is the parallel I keep returning to. In 2020 I audited the contracts of a high-yield farming protocol and found a reentrancy path that could have drained somewhere in the range of five million dollars. The community was celebrating yields at the time. I published a piece arguing that trustless finance without social consensus is not trustless, it is merely unattended, and it made me unpopular with people who were making money.

The thing I did not fully articulate then, and can articulate now, is what the yields actually measured. Television liquidity, enormous, headline-grabbing — and a substantial fraction of it was subsidized into existence by the incentive program. The number on the dashboard was not a measurement of interest. It was a measurement of subsidy. Remove the subsidy and the number does not decline. It vanishes, and it takes the appearance of legitimacy with it.

Content aggregation runs on the same mechanism. The volume is subsidized. Not by token emissions, by advertising inventory economics and by the zero marginal cost of syndication. The apparent editorial breadth is a subsidy artifact exactly the way the apparent user interest was. Take away the subsidy and you do not get less coverage. You get a much smaller number of items that someone actually chose to publish.

In both cases, the headline metric measured the incentive, not the thing the metric claimed to measure. That is why I do not trust any number whose production is subsidized, whether the subsidy is denominated in tokens or in pageviews.

The regulatory layer, briefly

There is one more structural piece, and it explains a specific pattern in coverage that I have tracked for several years.

Watch the volume of licensing-related coverage attached to a jurisdiction, and watch the timing. Coverage spikes are not correlated with regulatory events as strongly as they are correlated with competitive positioning between financial centers. When one jurisdiction launches a licensing framework, the coverage does not track the framework's technical content — it tracks which other jurisdiction the framework is competing against. The framing is almost always "who wins the hub." Rarely "what does this regime actually permit, and what does it prohibit."

This is the same failure as the genre tag. A category gets assigned according to the schema's incentive rather than the subject's content. The subject is a rule. The category is a race. And the reader absorbs the category because the category is what the headline carries.

The reason this matters for provenance is that it demonstrates the failure is not confined to lightly-staffed aggregators. It happens at the top of the market, in outlets with editors and legal review, because the incentive lives there too. Provenance problems are not a quality problem. They are an incentive problem that presents as a quality problem.

What a working verification layer would actually look like

I have been circling a design. Let me state it plainly, and let me state its cost.

First, signed everything. Not on-chain — signed at the point of authorship, with keys the author controls, using the same discipline developers have used for twenty years without ever needing a token. Signed commits are the closest existing analogue to a functioning provenance regime, and the reason they work is instructive: the signature verifies who wrote the line, the history makes attribution auditable, and revocation is a new commit rather than a court order. Nobody monetized it. Nobody needed to. It became infrastructure.

Second, transparency logs. Certificate transparency showed that an append-only public log can make a whole class of fraud structurally visible — not prevent it, make it visible. A logging layer for content origin would do the same. Not a truth layer. A visibility layer.

Third, first-class correction history. The absence of a modification timestamp in my misfiled item is not a small omission. It is the entire problem compressed into one field. A publication that records and publishes corrections is making a falsifiable claim about itself. A publication that does not is asking to be trusted on the strength of nothing.

Fourth, honest labels. "This item came from a partner feed. No editor reviewed it." Fourteen words. Free. Almost never present.

The cost is that someone has to pay. Signing keys have to be managed. Logs have to be operated. Corrections have to be written, and writing a correction takes a person, and people cost money. There is no revenue model for verification, which is why it is always the first line item to disappear when quarterly numbers get tight. This is the pragmatism test, and verification fails it, and that is the honest reason nothing has changed.

Contrarian

Here is where I want to push against my own instinct, because the comfortable version of this essay — the one where better cryptography and better standards fix the feed — is wrong.

The misfiled esports brief is not the scandal. The scandal is that we keep expecting the platform to solve a verification problem that belongs to the reader, and we keep being disappointed when the platform optimizes for what it is paid to optimize for. Every proposal I have seen in this space, including my own, asks the wrong party to bear the cost. Platforms will not. They are compensated for volume. Verification reduces volume. The economics are not close.

The second thing, and the more uncomfortable one: provenance is not truth. A perfectly signed piece of content can be perfectly wrong. A key proves control. A timestamp proves sequence. A log proves that a record was written. None of it proves the record is accurate, and any system that implies otherwise is building a lie with better tooling. I have watched crypto people make this error repeatedly — assuming that because a claim is on-chain, it is verified. The chain verifies the ledger. It verifies nothing about the world.

The third thing is the blind spot I find hardest to write about, because it is mine as much as anyone's. The people loudest about verification in this industry are frequently the least willing to sign their own work. Anonymous accounts demanding transparency from institutions. Pseudonymous analysts demanding provenance from publishers. Threads with no author entity demanding that media disclose its funding. I have done versions of this. The asymmetry is not hypocrisy, exactly — anonymity is legitimate and sometimes necessary — but it does mean the demand for verifiability is coming from people who have structured their own participation to be unverifiable, and that is not a stable foundation for a norm.

The last piece of the contrarian case is the perverse incentive that any labeling regime creates. Add a credibility badge and you have created a market for credibility badges. Two tiers become three, the third is fake, the fake is cheaper than the real one, and the label loses meaning within two quarters. I have seen this happen to ratings, to audits, to exchange proofs of reserves. Every verification mechanism becomes a target the moment it becomes a signal.

Takeaway

So the question I am left with is not how to verify content. It is who pays for it, and whether anyone actually wants it.

Here is what I suspect, and I would rather say it as a suspicion than a conclusion. Most readers do not want verification. They want corroboration — of a prior, a position, a trade. Verification is a cost they will not bear and a service they will not buy, and the market has correctly priced that, which is why the feed looks the way it looks. The scarce asset in the next cycle will not be blockspace, and it will not be intelligence. It will be the small number of readers who can tell the difference between a claim and a record — and the smaller number of writers willing to put their name on one.

Trust the protocol, not the pitch. And when there is no protocol — when the whole thing is just a feed with a taxonomy and no record of who touched it — the only honest move is to notice the missing field.

Code doesn't care about your intentions. Neither does the pipe. That is why I keep auditing it.