Trending Radar, in Detail

The signals overview covers when to use the radar rather than the briefing. This is what it actually does.

The radar answers one question: have several of my sources independently decided the same thing matters? Not what is popular on the internet — what is converging in the feed you chose.

The window

Forty-eight hours, split into two halves. The recent half is what is happening; the earlier half is the baseline it is measured against. That split is where "accelerating" comes from — a topic mentioned twice yesterday and eight times today is a different signal from one mentioned five times each day.

Up to 500 articles from your sources enter the clustering. That number sets the cost, not the window length, so a very busy feed sees the most recent 500 rather than a longer history.

The one hard gate

At least two distinct sources. Nothing else excludes a topic.

One outlet publishing six times about its own story is not convergence, and no amount of volume turns it into one. This is the only threshold that still throws anything away, and it is the premise of the whole feature.

What the score is made of

Every topic that clears the gate gets a score from 0 to 100, and the list is that ranking:

Component	Weight	What it measures
Breadth	40%	How many distinct sources
Volume	25%	How many mentions
Acceleration	25%	How much more than the baseline half
Freshness	10%	How recently it peaked

Breadth carries the most because it is the point of the feature. Freshness is a tiebreaker at 10% — enough to float this morning's story over one that peaked at the far edge of the window, not enough to rank anything on newness alone.

Each component saturates. A story in twelve outlets is not twice the story it is in six, and a ×40 acceleration off a single prior mention is an artefact of a small denominator rather than forty times the signal. So each input runs on a log curve with diminishing returns — breadth levels off around six sources, volume around ten mentions, acceleration around ×8.

Why a score rather than a flag

The radar used to mark topics "heating" with a yes/no test: three or more mentions and two or more sources and (new or accelerating ×2). Three thresholds ANDed together.

The binding constraint was almost always the mention count, and the case it excluded turned out to be the cleanest signal the radar has: two outlets, one article each, same story, nobody covering it the day before. Exactly the moment worth catching, thrown out for being small.

A flag also destroys the margin. A reader looking at "1 heating · 5 tracked" had no way to know whether the other five were noise or 2 mentions / 2 sources / ×4. The score keeps that information — nothing is excluded for being merely close, and the order tells you how close.

So rank by the score. The heating label is true for everything the radar returns, because clearing the source gate is what makes a cluster a radar topic at all.

The synthesis

Each topic can carry a paragraph synthesising what the sources are saying, rather than a list of headlines to open one at a time.

It generates once per day and is cached, so opening the radar repeatedly costs nothing extra.

The radar is free to run

It spends no AI credits. No model is involved in finding the topics — it is clustering and arithmetic over articles you already have. Open it as often as you like, on any plan that has it.

There is a 3-credit price listed for radar internally that is never charged; the synthesis is the only part that touches a model, and it is the daily cached one.

Pro plan and above. See plans.

If something looks wrong

The radar is empty — you likely have few sources, or they do not overlap. Two independent sources must cover the same story within 48 hours, which needs enough sources that any two of them can agree. Fifteen or twenty in one subject area is roughly where it starts working.

It shows things I do not care about — the radar reports what your sources converged on, not what you asked for. It deliberately ignores your interests: a story every source you trust picked up at once is worth seeing even when it matches nothing on your list. Change the input by changing your sources.

A story I expected is missing — check it came from two different sources. A scoop nobody else has covered yet cannot appear here by design; it will be in Recommended.

The scores all look similar — normal in a narrow feed, where every topic clears the gate the same way. The ordering still holds; the gaps are just small.