
Part 2 of 8 · Previously: a news app that has to explain itself · Next: Bayes’ theorem (coming soon)
By early September IntelliBrief could write a good briefing. That was also the problem. Everything it did was a function of today: the briefing, the topic digests, the editor’s note. Anyone with the same feeds and a similar prompt could produce something similar, which is why every comparison with other news apps ended in the same place.
What a competitor cannot reproduce is time. So on 7 September I started building what I call the constellation.
The idea
Every story IntelliBrief follows mentions people, places, organisations and things. The app was already identifying those names in every article (to match them against what a reader follows) and already recognising when today’s articles continue yesterday’s story. Then it threw both away at the end of the day.
The constellation keeps them. For each story it stores a connection to each name in it, with a strength (how central the name is to the story) and a confidence (how sure the app is that it identified the name correctly). Those connections are kept for 180 days.
From that one table you can read two graphs:
- A story’s graph: the names in this story, and which other stories share them.
- The day’s constellation: the names that mattered today, how they connect, and which are new since you last looked.
Story-to-story links are never stored. They’re derived when you ask for them, from the names two stories share. Storing every pair of related stories would have created a second graph that grows with the square of the number of stories. Over 180 days that design mistake would have finished the feature.
Because most of the bookkeeping already existed, the first version landed in two days. Everything since has been about making the lines mean something.
Problem 1: a line looks like a claim
The first review of the feature, on 9 September, put it bluntly: the graph is co-occurrence, not relation. Two stories share a name. That’s all a line says. But a thick line between a chip maker and a story about export controls looks like an assertion that one caused the other.
The code said this honestly in its comments and API documentation. Readers don’t read either. So the honesty had to move into the interface. The day’s constellation now has a How to read this key that explains every mark on the map: size, solid and dotted lines, arrows, glow, fading, and the caveat that a link means shared reporting, not a proven relationship. A Map / List switch shows the same names as a readable list, which is also where screen-reader users start.
Problem 2: everything connects through the United States
A naive graph of news collapses into a few giant hubs. “United States”, “Europe” and “energy” appear in so many stories that connecting through them carries almost no information.
The fix is old and good: weight each name by how rare it is across the stories the app has kept, the inverse-document-frequency idea behind search engines. A commodity trader that appears in three stories is a far more interesting link than a country that appears in three thousand. The scoring is explainable by construction: every term is a named quantity a person can argue with, and the app can print the reasoning behind any connection.
The same thinking led to a rule about identity. Related names are not the same name. England and the United Kingdom are distinct places. A head of state and their country need a real relationship in the evidence, not a merge in the vocabulary. Calendar words, generic nouns and bylines are kept out of the constellation altogether.
Problem 3: the graph only went up
The September review found what it called the single most important flaw: three separate mechanisms pushed every connection’s strength up and nothing ever pushed one down. A graph like that drifts towards everything being strongly connected to everything. Now a connection loses a fifth of its strength, and of its evidence, each time the story is updated without that name, so it can fall as well as rise. The story graph only shows connections above a floor of strength 0.20 and confidence 0.35.
Problem 4: a name in a story isn’t always what the story is about
This one came up today, 19 October, and it’s a good example of the work involved.
A reader noticed the word “Journey” on the day’s constellation. It came from a games-industry story: a new studio had raised $70 million, and its founders were described as “designers associated with What Remains of Edith Finch and Journey“. Journey is a 2012 video game. The story wasn’t about it at all. The name was a credential, a way of saying who the founders are.
Fixes followed…
Problem 5: gaps are not zeros
When a reader picks a name on the constellation, the app shows that name’s last two weeks of coverage in one sentence. If the collector didn’t run one day, that day is shown as a gap, never as a zero. The difference sounds pedantic. It isn’t: a zero claims “nothing was written”, a gap admits “we didn’t look”. This became the most important rule in the statistics work described in part 3.
Since 24 October the constellation also marks the names standing out today, with their numbers beside them…

Why it matters
The constellation is the first part of IntelliBrief with a real moat. Its value is the months of connections behind it, and those can’t be copied, bought or generated on demand. A competitor starting today is six months behind and stays six months behind.
It also changed what the rest of the app could do. With a memory of names and stories over time, it becomes possible to ask:
- is this name more prominent than its own normal level? (part 3, Bayes)
- which stories are episodes of the same longer development? (part 4, narrative threads)
- what is genuinely outside this subject’s pattern? (part 5)
- how widely is this covered beyond my own feeds? (part 6)
What I’d tell someone building something similar
- Store the cheap, primary facts (this story mentions this name, this strongly), and derive the expensive relationships when you read them.
- Penalise common names from day one. Every graph of text turns into hubs otherwise.
- Decide early what a line means, and put that meaning in the interface, not only in your documentation.
- Keep a way for evidence to go down as well as up.
Postscript, 28 October: a graph’s memory is not a reader’s memory
Keeping connections does not tell us what a reader previously saw. The new local continuity views add an explicit remembered theme comparison, not a belief inferred from graph navigation. The graph remains co-occurrence; source-linked reporting around a private idea is a different relationship. Part 7 follows that distinction into Library.
Next in this series
Part 3: Bayes’ theorem. A Bayesian model rejected by its own test data, the one that shipped, and accuracy gates that decide what readers see.
Try it
IntelliBrief is in beta and free while it is. The web version opens at intellibrief.sazalabs.com, and the iPhone and iPad app is on TestFlight.
Without an account you see only a small sample edition. Sign in to see what is really being talked about: the full day’s edition, stories with their history and themes, the subjects you choose to follow, and Tell me more on the stories you want to dig into.
Saeid Zarghani builds IntelliBrief.

One thought on “IntelliBrief, part 2: the constellation, a graph that remembers”