How IR Lens thinks

The Method

← IR Lens

Today, news is structurally biased toward events of the day, with a continuous, 24 hour news cycle. Large movements in global geopolitics, think the rise of a new empire, are affected by these events.

Figuring out how events fit in their respective bigger stories is difficult to do. IR Lens aims to leverage the first part to construct a clear picture of the second. We built a multi-level data model that borrows from research methods used by historians and intelligence agencies.

Storylines

IR Lens follows a select few important developments in global geopolitics: Storylines. Each storyline has a dossier that tracks developments, goals and movements. Every week, current events are measured against the storyline dossier. Will these impact the story, or not?

Each Storyline dossier contains the following four parts: - the Frame | long term, human driven. - the Estimate | Medium term, driven by monthly re-synthezised from the full record, previous Estimates, the Arc. - The Open Questions | Medium term, driven by the Estimate & the Record, story threads to keep an eye on. Updated weekly based on the Record. - The Record | Short term, driven by current events.

New storylines will be added in the future and current storylines might be archived if they’re no longer relevant.

Three layers of time and data

The French historian Fernand Braudel argued that history moves at three speeds at once. On the surface, histoire événementielle - the history of events: battles, summits, decisions, what he called the froth on the waves. Beneath it, the conjoncture: medium-term phases and cycles that give events their direction. And at the bottom, the longue durée: slow structures — geography, systems, mentalities — that barely move, yet decide what events can mean at all.

The IR Lens data model is based on these three layers:

Reports and estimates

Intelligence agencies learned the same lesson in different words. Sherman Kent, who founded analysis as a discipline at the CIA, insisted on separating current intelligence — this week's reporting — from estimative intelligence: the considered judgment of where things stand and where they are heading.

In IR Lens the Sunday email is current intelligence; the dossier is the estimate. The digest also inherits Kent's discipline of probability: we try to hedge estimates with numbers and what is known is kept distinct from what is inferred.

From raw news to a week view & an updated dossier

Let’s follow one week to see how the model goes from raw input to an updated dossier and weekly digest.

Our (current) eighteen sources produced 272 news items; a keyword sieve discarded half as not relevant; clustering collapsed the rest into 40 events. Every event was scored against every storyline dossier. In the end, three analyses were written.

What makes this a storyline system rather than a summarizer is the pair of red arrows on the right. The weekly analysis reads the relevant storyline's dossier, from its last ten entries to its open questions and uses it for input. After writing the weekly digest, it updates and extends the dossier. Each week's output becomes part of the next week's input.

THE STORYLINE DOSSIER the frame the struggle it tracks, its subjects, its theoretical lenses — hand-written, changes only by deliberate decision open questions at most ten, resolved and replaced as the arc moves the record one entry per analyzed event — grows every week, and is the continuity every later analysis is anchored on temperature series weekly tension reading — drives the gauges the estimate monthly re-synthesis of the full record — human-approved, then anchors the record sources 18 feeds gather raw reporting 272 items sieve kept 126 · no model cluster events 40 the frame's subjects gate the sieve score every event against every storyline, 0–10 a strict rubric that rewards change, not volume the frame temperature computed from the scores, not prompted from a model selection top pairs · at most six a week wildcard strong event, no home — a storyline candidate analysis the next entry in an ongoing record, never a standalone essay reads: last ten entries · open questions writes: appends its entry · updates the questions appends the weekly reading
One week's build, with a real week's counts. Black arrows run once per week; the oxblood pair is the compounding loop between analysis and record.

Monthly re-synthesization

Once a month we re-synthesize every storyline. We do this for a reason. The psychologist of intelligence analysis Richards Heuer documented a failure mode every news reader knows from the inside: analysts who follow a story incrementally, week by week, become worse at seeing the current picture than someone who arrives fresh — each new increment gets assimilated into the frame that already exists.

So the review deliberately starts over: it re-reads every analysis in the record, forms a fresh synthesis, and then compares itself against the previous dossier.

the full record every entry, not just ten re-read draft estimate arc · thresholds · scoreboard human review read · edit · approve the standing picture the approved consolidated record becomes the anchor the weekly analyses are written against · refreshes the open questions
The monthly loop. The dashed box is the human gate: no estimate becomes the standing picture without being read, edited, and explicitly approved.

who or what does the analysis?

The drafting at every layer is done by a language model; that is what makes a one-person operation like this possible. But that is not to say a human isn’t involved. The weekly digest runs autonomously, but I, a human, edit and check before sending. Same for the monthly dossier: every draft goes past me. I read it, edit it, and explicitly approve it before it becomes the standing picture. And in the background, every prediction the system makes is tracked and scored in the weekly analysis as fired, expired, or still pending. Depending on outcomes, I can tweak and adjust the system.