How Stories Travel / Method

How Chains Are Matched

A chain links appearances of one claim across languages and dates the first located appearance in the outlets MEF monitors. It describes text and timing. It does not say who wrote first, why, or whether outlets worked together.

Matching

Each item is a headline plus its first 40 words, in the original language and, where the press monitor produced one, a machine English rendering. A multilingual sentence model (paraphrase-multilingual-MiniLM-L12-v2) places every text in one vector space. Two items from different outlets link when their best similarity, across original and English texts, reaches 0.82 and their feed timestamps sit within 3 days. Figures that appear in both texts must overlap, or the score drops by 0.08. Candidate pairs scoring 0.74 or more were sent to a language model, which compares both texts in the original language and vetoes a pair that reports a different statement. In the current candidate chains, 2,412 of 2,443 cross-outlet pairs carry a model check; an unchecked pair is not vetoed, and two items from the same outlet are never compared directly. The fluent reviewer checks every step before publication. Chains form by complete linkage: every pair in a chain must clear 0.77, and a chain holds at most 25 items.

Measured Accuracy

The test set is 80 pairs drawn from monitor data (40 matches, 40 near misses on the same topic), labelled by the build process from the original-language text and a machine English rendering, not yet by fluent staff. The threshold was chosen on these same pairs, so the figures at it are in-sample: precision 100% (33 of 33 links correct) and recall 82% on the calibration set. A 5-fold held-out check, which picks the threshold without the pairs it scores, gives precision 90% (35 of 39) and recall 88%. Every labelled pair carries a language-model check; not every production pair does (see Matching). Fluent-staff relabelling is pending. Labelled set and scores: qa/labels.jsonl, qa/eval.json.

Dating the First Located Appearance

The date is the outlet's own feed timestamp, converted to UTC. Where a feed gives a date with no time, the step shows the date alone, a date-only item never takes step 1 from a timed item on the same day, and a chain whose order depends on a date-only item is held for review. Where the Internet Archive holds a capture, its first capture time is shown beside it as separate evidence; it never replaces the feed timestamp. "First located appearance" means the earliest item found in MEF-monitored sources as of October 17, 2026. An earlier appearance elsewhere is a correction, listed below.

Sources Searched

Publication Standard

1. A fluent reviewer checks every step in its original language, not the machine rendering.
2. Every step is archived at publication time.
3. Any outlet outside state or state-directed media receives a request for comment with 72 hours to reply; replies are published with the chain.
4. A named editor signs off.
5. Telegram, X and broadcast archives are checked by an analyst for an earlier appearance.
6. Quotations stay under 50 words per step; each step links the original.
7. Selection follows the same rule for every outlet and country, including Israeli and US sources. Chains that centrally concern a US federal candidate are held through Election Day.

Corrections

• None so far.

Send an earlier source or a correction to info@meforum.org. Each accepted correction is dated here and the chain is versioned.

Cite This Section

Middle East Forum, "How Stories Travel," MEF Press Readout, accessed October 17, 2026, https://readout.meforum.org/stories/.