How the paper computes its numbers
Every figure the paper prints about itself comes from code that reads the archive at build time: the count of sources under a story, the forecast record and the World Desk index. This page explains each of them and shows where each one falls short.
The Record
Each story ends with the Record: one row per artifact the desk retrieved, numbered E1, E2 and onward, and cited in the body by that number. A row names the outlet, gives the passage the story rests on, and opens to a source note with the excerpt, the link, the retrieval time and a line saying how the excerpt was obtained.
- Verbatim passages. The research capture names the sentence it relies on. A checker then re-reads the live page and files the excerpt only if the page carries those words in that order. Curly quotes, dashes and punctuation are ignored, but every word has to match. A longer passage is named by its first and last words. The checker copies the page’s own text between them, up to five sentences and 1,200 characters, and refuses anything longer instead of cutting it.
- Speakers kept. When a quote names its speaker only as “he said”, the checker adds the page’s previous sentence, where the name usually is.
- Page furniture removed. The passage is first found on the page exactly as published. Advertising slugs and “Also read” link boxes are then cut from it. A passage that still carries anything like furniture, such as a “Related:” line, a newsletter prompt or an image credit, is refused rather than trimmed, and so is a filing whose excerpt carries one.
- Attributed reporting. Some pages cannot be opened by the checker: paywalls, blocked requests, PDFs and timeouts. Those rows are marked as attributed reporting. They carry no excerpt and are never quoted. They can support “the outlet reported that…” with the link, and never as a story’s only support.
- Sentences the page does not carry are dropped. If the page was read and the quoted sentence is not on it, the row does not reach the Record.
Counting sources
Ten rows in a Record are not ten sources. Several rows often quote one page, and several outlets often relay one government statement. Each Record header therefore carries three counts.
- References
- Rows in the Record, E1 to the last.
- Outlets
- Distinct publishers among those rows: one per web domain (bbc.co.uk and bbc.com are one), and one per account on a social network. The paper’s own computations are rows but not outlets.
- Independent
- Outlets left after relays are merged. Only an outlet with at least one excerpt the checker found on its page counts. Attributed reporting and rows without a captured excerpt never count. Two outlets are merged when an excerpt from one credits the other (“RBI Governor Sanjay Malhotra said” in the Indian Express merges it into the Reserve Bank of India), when excerpts from both credit the same named speaker or agency (“Rumen Radev said”, “according to a Reuters report”), or when they share ten consecutive identical words, as a reprinted wire story does.
One relayed excerpt merges a whole outlet, so the count leans low. It is still pattern matching on the excerpts, not proof. A relay worded in a way the patterns miss is not merged, and neither is a briefing the excerpts do not name: two outlets repeating it count as two. The number means distinct publishers the Record does not show to be repeating each other. It does not prove that each outlet reported the story separately.
Forecasts
A call is a yes-or-no question about the world with its deadline written into the wording, in the form “Yes if the council passes the budget by 30 November”. The forecaster files a probability that the answer is yes. That probability is the posterior printed in the Forecast Ledger. At 0.5 or above the paper’s call is YES, and below 0.5 it is NO.
Once published, that posterior becomes the call’s prior: the ledger keeps the probability printed with the call and scores the call on it, even if a later story argues a different number. A call stays on the ledger from edition to edition until Ledger, the settlement clerk, settles it. Every call is in one of five states:
- Not yet due
- The deadline in the call’s wording has not arrived.
- Due, awaiting verification
- The deadline has passed, or the call names no date, and Ledger has not settled it. The row prints Ledger’s note on what was checked and why the call is still unresolved.
- Settled hit · settled miss
- Settled against a record Ledger can name, which is quoted in the note. A hit means the call came true: a YES call whose event happened, or a NO call whose event did not.
- Cancelled
- The event can no longer happen as worded, or the call was withdrawn. The reason is given. A cancelled call is scored neither way.
Track record
As of the 11 October 2026 edition, the archive holds 59 settled calls: 35 hits and 24 misses. Another 5 are open: 5 not yet due and 0 due and awaiting verification.
The Brier score is the mean of (probability − outcome)², where the outcome is 1 if the event happened and 0 if it did not. Zero is perfect, and lower is better. A forecaster who says 0.5 every time scores exactly 0.25. The base-rate line uses a single probability for every call: the share of events that did happen (58%). That share is only known afterwards, so it is a stricter yardstick than the coin flip. The paper currently scores better than both.
Calibration checks whether calls made at a given confidence come true about that often. Each call is grouped by the confidence the paper put on its own call: a 0.62 probability of yes and a 0.38 probability of yes are both 0.62 confidence, the first in YES and the second in NO.
A well-calibrated row reads about the same in Stated and Came true. With a handful of calls per row, the gap is mostly noise, so read the table as a direction rather than a verdict. Calls published before 29 September 2026 were scored when settled but were not carried forward while open. Their unsettled remainder is not in these counts.
The World Desk
The escalation index on the front page (0.46, steady in this edition) measures market and energy stress linked to conflict. It does not measure conflict. It is the share of a fixed list of public price series that stand at or above a stress threshold, with each series weighted by a severity set in advance:
index = Σ severity of series past threshold ÷ Σ severity of all series
Each threshold is that series’ own 80th percentile over the five years before the list was frozen on 6 September 2026. The thresholds are not re-tuned to produce a number. The scale runs from 0, with no series past its threshold, to 1, with every series past it. With a total severity of 13, the index moves in steps of 1/13, so 0.46 means 6 of 13 severity points are past their thresholds. “Steady” means the reading moved less than 0.005 from the last one, “rising” or “easing” gives the direction otherwise. A series whose latest data is missing or too old stops the reading for the day rather than being guessed. A war that moves no market leaves the index unchanged. That is a limit of the measure, and it is printed here for that reason.
The conflict counts beside it (2 open, 3 on watch) come from a fixed list of thirteen standing flashpoints, such as the Strait of Hormuz, the Ukraine–Russia front and the Taiwan Strait. A place is open when at least two of the day’s stories, each with a source link, mention it. It is on watch when one story does, or when it was open in one of the previous seven editions and drew no story today. Matching is by keyword, not by understanding, so a passing mention can count.
The numbered markers on the globe and in the flashpoint index beside it are an editorial choice: the places behind the day’s stories, keyed to their coverage. They are not scored, and a probability printed beside a marker is not part of the track record above.