CASE STUDY — Market Moves · RESEARCH DESKS / MARKET INTELLIGENCE · NEWS INTAKE & ANALYST ALERTING

Market Moves: 100,000 Articles a Day, Down to the 34 That Clear the Gate.

Market Moves promises one thing: fewer alerts, zero false ones. A research desk drowns in news it has no reason to believe, and the constraint was never coverage. It was that one unconfirmed story, read at the wrong moment, ends up as a footnote in a research note somebody downstream relies on.

BY Sarah, Technical Editor | September 7, 2026 · SEPTEMBER 7, 2026

99.97%

OF WHAT ARRIVES IS DELIBERATELY WITHHELD

~100,000 ARTICLES A DAY FROM ~5,000 SOURCES, READ AND CLUSTERED
ONE DAY END TO END — 61,400 DECISION-RELEVANT · 2,140 EVENT CLUSTERS · 312 CORROBORATED · 34 PUSHED
THRESHOLD SET BY THE ANALYST — 2 TO 5 INDEPENDENT SOURCES, 30 / 60 / 120-MINUTE WINDOW

01 — THE PROBLEM

A hundred thousand a day, and three fixes that don't hold

A hundred thousand articles a day. Five thousand sources. One analyst. One morning. A number they will have to defend to people who never saw the source. The obvious fixes fail in different places. Keyword alerts fire on the word rather than the event, so five differently-worded reports of one incident arrive as five alerts while a reposted rumour arrives as news. Ranking sorts what it shows you and still shows you everything. Two more analysts read twice as much and still guess. Every other feed optimises for showing you more, and more is the wrong axis when being wrong is the expensive outcome.

02 — INSIDE THE GATE

You set the threshold. The feed obeys

Nothing ships until independent outlets report the same event inside a window the analyst picks: two to five sources, and 30, 60 or 120 minutes. The rule holds on a quiet day as well as a loud one, and we do not relax it to fill a feed. The window is not a delay anyone sits through. It sets how far apart those independent reports may fall and still count as one corroborated event, so widening it admits slow-building stories rather than adding latency. Independence is part of the count. Two outlets under one owner are counted once, because corroboration that resolves back to a single newsroom is not corroboration. When nothing clears the bar, nothing is sent, and the app says so on screen instead of padding the day. What does clear it arrives as a push on a mobile client, carrying the sources, the spread between them, and the thread of updates the event belongs to.

03 — THE METHOD

The index isn't for search. It makes corroboration countable

Counting independent outlets sounds trivial until you try it. The example we use to explain it is a chip-plant shutdown written up as five headlines with no shared keywords: "Quake forces shutdown of Fab 4, company confirms" beside "Wafer supply at risk as plant goes offline". Vector similarity collapses those five into one countable cluster: five independent outlets, forty-one minutes apart, one event. The count, not the text, is what clears the gate. The stack underneath is ordinary: a 384-dimensional FAISS index, MongoDB, Python services on AWS, a React Native client. None of that is where the product lives. The decision rule is.

Most teams use embeddings to retrieve. We use them to measure agreement.

04 — WHAT IT REFUSES

The hard part is refusing, and refusing does not demo well

One day, end to end: about 100,000 articles in, 61,400 judged decision-relevant, 2,140 event clusters, 312 corroborated, 34 pushed. Sports and entertainment are dropped at ingest. Single-source stories never reach a device. The survivors thread into per-topic timelines — monetary policy, supply chains, earnings, logistics, commodities, regulation. It earns its keep by what it refuses to send, and that is the part that is hard to sell. A product whose good day is a quiet one does not demo well. We are not publishing revenue, customer counts or an accuracy figure here, and we would rather leave the gap visible than fill it. So here is the offer instead. You tell us what your research covers. We set the threshold with your analysts, replay seven days of intake, and hand back the feed with everything we refused to send beside it. You keep the rejections.

Need a feed that
earns its interruptions?

Talk to the team Agentic AI service →

FAQ

Questions about this product

If you withhold 99.97% of what you read, how do I know you aren't withholding something I needed?

Because you get to see the refusals. In the seven-day replay we hand back the feed and everything that did not clear the gate alongside it, so you can check the rejections against your own coverage before you commit to anything. Day to day the bar is yours, and loosening it is one setting. The trade is visible when you do: more items, weaker corroboration.

Doesn't waiting for corroboration mean my analysts get the news late?

Later than the first post, yes, and that is the design. But the window is not a delay you sit through. It is how far apart the independent reports may fall and still count as one corroborated event. Set it to 30 minutes and you get only what the market converged on fast; set it to 120 and slower-building stories qualify too. Either way nothing single-sourced ships. If being first to an unconfirmed story is the job, this is the wrong product and we would rather say so now.

What stops five outlets republishing the same wire from counting as five sources?

Independence is part of the count, not a filter applied to it. Two outlets under one owner are counted once, and the number attached to an item is independent outlets, not articles.

"Zero false ones" is a strong claim. What is that actually a promise about?

It is a promise about the gate, not about the world. Nothing single-sourced reaches a device, and nothing reaches it before it clears the threshold you set. We are not claiming an outlet can never be wrong, or that several independent newsrooms can never be wrong together — the count measures agreement, and agreement is not truth. What we publish is the rule and the funnel it produced on one day, roughly 100,000 in and 34 out, so the claim can be checked rather than taken on faith.