At 06:14 Geneva time my ingestion script dropped a headline into the geopolitics bucket: NRCC targets Trone Garriott's record in Iowa's 3rd District. The publisher was a crypto outlet. No timestamp on the page. No byline. No poll number, no ad-spend figure, no rating from any of the three services that rate House races, and no explanation of what the word "record" actually refers to. A headline, a lede, and the syndication footer that tells you the item was pulled from somewhere else and republished.
I have run some version of that pipeline since 2017, when my triangular arbitrage scripts needed headlines faster than I could read them. It has paid for itself many times over. Lately it pays me mostly in noise. This item is the cleanest specimen of the quarter: a crypto publication running a US congressional race brief that has nothing to do with crypto, nothing to do with markets, and nothing inside it you can verify. The ledger doesn't care that a story was published. It cares what the story contains. This one contains a sentence.
Honest framing first: this is not a story about Iowa. The 3rd District is a suburban Des Moines seat, the kind of swing district that decides House majorities, and it will be fought over by people with real money and real data. None of that is in the item. What is in the item is a supply-chain artifact, and the artifact is the thing worth auditing.
Crypto media economics explain it. Display CPMs on aggregator traffic have compressed for years. The marginal cost of producing a page has gone to roughly zero โ an RSS pull, a template, a republish. Verticals that once competed on original reporting now compete on inventory: how many pages can be filled per hour with something technically true. So a publication whose entire value proposition is crypto-specific runs congressional race briefs, because the feed had a slot and the slot was empty.
The syndication footer is the tell. Items republished through aggregator pipes carry a standard attribution line, and that line is a forensic marker: no editor assigned the story, no reporter called a source, no desk verified a number. File it under source grading. The second-order problem is worse than the first. Run a corpus of these items through a summarizer and it will manufacture significance out of volume โ three hundred empty items become a trend, because the model was instructed to summarize, not to reject. I have watched an assistant produce a confident paragraph about rising political risk in the American Midwest from a bucket of exactly this material.
A useful version of this brief would have taken four fields: publication date, election cycle, current rating from a nonpartisan handicapper, and any 48-hour independent expenditure filings naming the committee. That is a ten-minute pull for anyone with a terminal. Its absence is not an accident of brevity; it is the difference between a data point and a placeholder.
I run three checks on every ingested item before it reaches my attention. All three are automatable. This one fails all three.
Check 1 โ platform-content consistency. A crypto vertical covering an Iowa House primary is a categorical mismatch, and mismatch is a hard, machine-detectable feature. You don't parse the article; you count. For each source, track the share of items failing an on-topic test over a rolling 30-day window. Three off-topic items in thirty days is a downweight signal. Ten is a delisting signal. The threshold matters less than the fact that you measure it continuously, because contamination is monotonic โ feeds get dirtier over time, not cleaner.
Check 2 โ information density. Count the verifiable numerals: poll margins, dollar figures, dates, vote counts, filing identifiers. This item has zero. Not one number appears anywhere in the piece. A news item with no numbers is not news; it is an announcement that something occurred. Announcements belong in a different queue from data, and merging the two queues is how a monitoring system quietly turns into astrology.
Check 3 โ timestamp anchoring. This is the fatal one. No publication date, no identified cycle. Was this written inside the 2024 cycle or ahead of 2026? Without an anchor you cannot place the event on a timeline, and without a timeline no statement about momentum or contest means anything. An undated item is not low-confidence information. It is information with the independent variable removed.
The cost is not abstract. Analyst attention is the scarcest input in any monitoring operation, mine included. Every false positive burns review time, and worse, it inflates the apparent base rate of geopolitical events in the feed, which degrades precision for every downstream alert. Alert fatigue is a precision failure wearing an operational costume. Volatility is just unpriced fear wearing a mask; feed noise is the same trick at lower stakes.
Score each item before routing it. I use a blunt composite: timestamp present, verifiable numerals capped at five, platform-content match, resolvable named entities. Anything under four points goes to a review queue, never to an alert queue. The effect on precision is not marginal. On my own feed, filtering by those four fields cut alert volume by roughly two-thirds and removed almost nothing I later needed. That ratio is the whole argument. The rejected material was not slightly less useful. It was inert.
Now the two things the item did give me.
Institutional metadata beats prose. The attacking committee is the National Republican Congressional Committee, whose charter is electing Republicans to the House. Therefore the target is a Democrat. That inference needs no lookup, no briefing, no source โ it falls out of the attacker's mandate. I extracted a verified fact from a piece of content containing no facts, purely by reading structure instead of sentences. If your pipeline parses text and discards publisher, committee, filing entity, and jurisdiction, you are throwing away the only reliable signal in low-quality items.
Template language works as a timestamp substitute. Targets-plus-record is fixed phrasing in American political reporting. It surfaces at one stage: opposition research has produced a deliverable and is testing or placing it. That tells you material exists and the people who paid for it consider it deployable. It does not tell you whether the material is a voting record, a business dealing, or a legal dispute โ and record is among the most elastic words in the political vocabulary. That ambiguity is the largest gap in the item, and it is the gap that decides severity: financial or legal material escalates, a legislative voting record is routine.
One contradiction worth flagging. The item calls the race tightly contested and says the attack may shift it, while supplying no baseline โ no prior margin, no current poll, no rating. A conditional claim with no baseline is not analysis. It is the grammatical form of analysis with the substance removed.
Crypto readers should not scroll past this, because the chain from a suburban House seat to asset prices is real, just heavily attenuated. District result feeds House majority, decided across roughly forty to sixty comparable seats. Majority feeds control of the NDAA agenda. Agenda control feeds appropriations scale and structure โ munitions replenishment, shipbuilding, nuclear modernization. Appropriations feed defense order books and European defense valuations. Four stages, each with a brutal decay factor. A single district is an input variable to a majority model, not an object of analysis. Anyone deriving a defense budget view from one House race is committing a level error, and the error compounds when the source item contains no numbers at all.
In 2022 I made most of my money shorting things whose failure was legible in advance โ Celsius, Voyager, LUNA โ because the leverage was printed in the contracts and the filings. Political money is legible in the same way, just slower. Independent expenditures hit the FEC within 48 hours of a buy; ad-tracking services publish weekly totals. If you want exposure to this chain, instrument the aggregate โ swing-district spending, handicapper rating migrations, the appropriations calendar โ not the individual race. That is a real dataset. The brief in my feed was not.
For crypto portfolios the link is weaker still and runs through the macro factor that prices everything else: liquidity, duration, fiscal trajectory. In 2024 I tracked twelve institutional addresses that accumulated roughly 45,000 BTC across the quarters before the ETF approvals and published a flow-based thesis. It held because the flow data was unambiguous and the political narrative around it was noise I could ignore. That ordering has not changed.
The prevailing worry about information quality is synthetic media โ deepfakes, fabricated quotes, generated articles. Real, but misdirected, because synthetic falsehood leaves provenance gaps and inconsistent entities, and a fact-check can catch it. The actual degradation is duller. It is true, bland, headline-level content stripped of every quantity that would make it useful. Non-information rather than misinformation. It passes every truth check, matches no known falsehood pattern, and clogs the pipe anyway.
The second blind spot is directional. Most traders assume the on-chain layer is the dirty one and journalism is the clean one. Right now it is the reverse. Wallet flows don't editorialize, don't syndicate, don't fill empty slots. They also don't explain themselves โ a transfer is a fact without a motive โ but they never lie about what happened at a block height. Silence is the only honest signal in the noise, and the chain is the closest thing to silence this industry has.
The next US cycle will push political spending into every feed that will accept it, crypto feeds included, and contamination rates will rise before they fall. Build the filters now: mandatory timestamp, minimum one verifiable numeral, platform-content consistency score, rolling off-topic rate per source. Risk isn't a variable you control after ingestion; it's the one you define before the first packet arrives. Arbitrage waits for no one, and neither should you.