On a Tuesday in the third week of a quarter I've stopped counting, a military conscription item appeared on Crypto Briefing. A visitor to Uman — an Israeli-Ukrainian, by the item's own description — had been swept into Ukraine's mobilization apparatus. Both factual claims carried the same tag: Source: none.
I didn't care about the story. I cared that my scraper caught it.
My pipeline flags domain-topic entropy, a cheap Shannon calculation over the topic distribution of every feed I ingest. Crypto Briefing had been drifting for weeks, but this was the first time a hard military-conscription item cleared my threshold. The carrier didn't match the cargo. That mismatch is a signal, and signals are what I trade on.
Crypto media has been mutating since 2022, and almost nobody has priced the mutation.
The mechanics are boring, which is why they're reliable. Crypto-native display advertising collapsed with the broader ad market; per-reader revenue on pure DeFi coverage fell hard enough that outlets needed volume from somewhere. The audience didn't vanish — it aged into macro. The same person who once read a Uniswap V3 tick analysis now also reads rate-path speculation, sovereign-debt chatter, and, apparently, Ukrainian mobilization notices. Serving that reader across a wider topic surface isn't corruption. It's arithmetic.
But arithmetic has second-order effects. When a feed broadens its topic surface, it must either expand verification capacity or dilute it. Expanding verification is expensive: stringers, desk editors, legal review. Diluting is free. So the ratio moves.
This matters to anyone holding a position, because crypto feeds are no longer only news for humans. They're inputs to sentiment models, LLM-based trading agents, and social-signal pipelines. A feed stops being a feed the moment its attribution density becomes a rumor. When the input degrades silently, every downstream model inherits the degradation without a warning light.
I've spent seventeen years watching crypto infrastructure and six of them inside on-chain data, most recently building anomaly detection on Dune. My last published work, "The Ghost in the Ledger," found that roughly 15% of volume tagged "organic" across a one-million-record sample was actually bot-generated coordination. Synthetic volume. It looked exactly like real volume until you measured the metadata.
This is the same problem one layer up. Synthetic narrative looks exactly like real narrative until you measure the metadata.
Over 180 days I ingested headlines and opening paragraphs from nine crypto-desk outlets — roughly 14,000 items. For each item I computed three things. The Attribution Index (AI-1) is binary: does the item name a primary source — a person, a filing, an on-chain address, an official statement, a wire credit? "Source: none" scores zero. Carrier-Topic Entropy (H) is the Shannon entropy of an outlet's topic distribution over a rolling thirty-day window; rising H means broadening. Correction Latency is the hours from publication to any appended correction or clarification.
The Attribution Index fell from 61% to 38% across the panel. Nine outlets, one direction, no exceptions — a twenty-three-point decline in six months.
The decline was not uniform, and that's the interesting part. Native-crypto items — protocol exploits, governance fights, exchange flows — held attribution near 70%. Items in the broadened categories, geopolitics and macro and defense, came in at 19%. The dilution is not happening everywhere. It is happening precisely where the feeds expanded.
Entropy confirms the shape. Median H rose in seven of nine outlets. The two that held flat had both added a dedicated verification desk during the sample window. I'm not willing to call that causal on n=2, but I've filed it under "revisit with more data," which is where I keep the things I suspect and can't yet prove.
Correction latency tells the same story from a different angle. Median latency was 31 hours for broadened-category items versus 6 hours for native-crypto items. Broadened items get corrected later, or not at all — about 12% received no correction even after a named source subsequently contradicted them.
Then there was the Uman item, which I tried to trace the way I trace everything. Follow the gas. Always. For an on-chain claim, that means following the transaction. For a traditional-finance claim, it means following the filing. For this item there was nothing to follow: no wallet, no contract, no venue, no filing, no named official, no wire credit. The item sat in a state that would have been structurally impossible in crypto coverage five years ago — a zero-on-chain-footprint claim published on a crypto domain.
That is the finding. Not that the item is false. I have no evidence it is false. The finding is that the infrastructure to publish an unfalsifiable item now exists inside the crypto information stack, and it is being used.
I ran the cross-check anyway, because that's the job. I pulled the address clusters associated with Ukrainian volunteer-funding flows on Ethereum and Tron over the same week, looking for any anomalous transfer pattern consistent with a mobilization event at that scale. Nothing. The on-chain surface was flat. That tells me the claim, whatever its truth value, does not touch the ledger — which is exactly why it can travel so far without friction. A wallet can be traced. A sentence cannot.
Here is the parallel that keeps me up. In "The Ghost in the Ledger" I could not tell synthetic volume from organic volume by looking at the volume. I could only tell by measuring the metadata — wallet clustering, funding provenance, timing regularity. A headline is the same shape. You cannot tell a sourced item from an unsourced one by reading the prose. Both are fluent. Both are confident. You can only tell by measuring the apparatus behind them.
Data Integrity Checks. I owe you my error bars, because a number without them is a rumor in a suit.
My sample is nine outlets. Nine is not a population, and I won't pretend it is. Binary attribution coding is lossy: an item citing "a person familiar with the matter" scores the same as one citing a named official, though their epistemic weight differs by an order of magnitude. Topic labels came from a classifier trained on 4,000 hand-labeled items, and my inter-rater agreement was 0.81 Cohen's kappa — decent, not clean, with perhaps one in five marginal items mislabeled. I cannot and do not claim the Uman item is fabricated. Source absence is not source falsity. Code is law; math is evidence. I measured metadata, and metadata is not truth. Confusing the two is the exact error this article exists to flag.
So what is the actionable read? Treat attribution density as a leading indicator, the way I treat gas-price divergence. When attribution falls, feed quality falls before anything visibly breaks. A sentiment model trained on a feed whose Attribution Index has crossed below roughly 40% is being trained substantially on noise. You do not need to know which items are false. You need to know the density is low, and size your confidence accordingly.
This is the same discipline as volatility. Volatility exposes leverage. When the market moves you don't ask who was risky — you already know, because the move told you. Attribution decay is the quiet precursor; the move is the reveal. By the time a feed publishes something false and it moves a price, you have already lost the window.
The lazy conclusion is that crypto media is broken. I will argue against my own data.
I ran the same pipeline on three traditional finance wires over the same window. Their attribution decline was real but smaller — roughly eight points against my twenty-three. The direction is identical. Crypto feeds are not uniquely degenerate; they are earlier on a curve the entire information economy is walking.
More uncomfortable still: maybe the broadening is correct. A trader in 2026 genuinely does need to price a sovereign-mobilization risk, because that risk reaches the tape through energy, defense, and rate channels. Refusing to cover it is not purity; it is blindness. The failure is not the topics. The failure is carrying the topics without carrying the verification apparatus the topics demand.
And here is where most of my colleagues go wrong. Correlation is not causation. I found attribution decay and topic broadening moving together. I did not prove that broadening caused the decay. Both could be downstream of a single cause — a shrinking per-reader revenue pool that punishes verification spend and rewards volume. I flagged it. I cannot yet resolve it. Anyone who tells you otherwise is selling a narrative, and you now know how to check the metadata on that.
Watch one number next quarter: the Attribution Index of your primary crypto feed. If it holds above 50%, you have an input. If it slides under 40%, you have a mood. The question is not whether crypto media can cover the world. It is whether it can afford to verify it — and whether you will notice the moment it stops.