What does the number measure?

How we score — a worked example

“Mazag Masr” = The pulse of Egyptian talk online, in a glass of tea. The number 51/100 is A reading of the tone of the public content we collected during the week of 26 Oct – 1 Nov. Not an opinion poll

• ✅ It measures: the tone of published public content (news, comments, trends, songs, films…) — positive, negative or neutral.
• ✅ And: what people talk about most (each area's share of the conversation).
• ❌ It does not measure how people themselves feel, nor the view of “all Egyptians”. People who post online are not a representative sample.

From collected to counted

Week of 26 Oct – 1 Nov: from everything collected to the weekly number

1. Collected from public sources

   Through the week of 26 Oct – 1 Nov, from news headlines, YouTube comments, Telegram channels, trending videos, Reddit posts, chart songs, search & X trends, box office and humour channels.

2. Tagged by AI

   Repeated or extra headlines are tagged from a sample only; every other source type is tagged in full.

3. About Egypt (relevance filter)

   Anything not about Egypt is dropped from the count.

4. Unique stories (repeats grouped)

   A story repeated across sources is grouped into one; re-posts stay in the count but down-weighted (never double-counted).

5. Items counted in the weekly number

   Any area whose material this week is under the minimum (“too little material”) (like transport this week) gets no number this week, and is left out of the national index.

This is data collected and tagged throughout the week, not a single day's snapshot. The live pulse reads more material through the day, but the weekly number is computed only from material that passed every filter above.

A real example from this week

One item, end to end

1. 1

   The item

   «تفاصيل أسعار الخضروات والفاكهة اليوم الأربعاء 28 أكتوبر 2026 في الأسواق»

   news headlines · published 2026-10-27

2. 2

   Tagging

   • Area: Cost of living & prices
   • Tone: -1 (negative) on a −2…+2 scale
   • Joke? no
   • Local or global? local

   Tagging is done by an AI model (Claude Haiku for headlines, Sonnet for comments) with instructions for Egyptian dialect and sarcasm — it can be wrong, which is what the consistency test below checks.

3. 3

   The weight

   0.5headline weight (edited media, not public voice) × 1/√5 = 0.447story repeated 5× × 1local = 0.224

   Comments and trends weigh 1, videos/songs/films 0.8, Telegram news 0.7, headlines 0.5. A story repeated many times has its weight split; global items count half.

4. 4

   The area number: Cost of living & prices

   D = 50 + 25 × Σ(w·s) ÷ (Σw + 6)

   50 + 25 × (-35.02) ÷ (52.44 + 6) = 35

   All the material in this area this week; ours is one item in it, weighted 0.224. The “+6” pulls thin areas towards 50 (neutral).

5. 5

   The weekly number

   Each area is weighted by its share of the talk (smoothed, capped at 20%). Our area weighs 5.3%, so its part of the weekly number = 35 × 0.0528 = 1.85 of 51.4 points. Because its number (35) is below neutral (50), this area pulls the weekly number down.

   Area	Number	Weight	Contribution
   General news	44.1	12.1%	5.35
   Trends, jokes & memes	66.4	10.3%	6.86
   Football & sport	62.1	10.3%	6.4
   Work, money & the dollar	50.2	8.0%	4.02
   Series & films	49.9	7.4%	3.71
   Songs & mahraganat	60.1	6.5%	3.9
   Cost of living & prices	35	5.3%	1.85
   and 13 more areas			…
   Sum = the weekly number		100%	51.4

Confidence labels

Five separate labels — each measures one thing

• Volume

  Items in the area: 80+ = large · 30–79 = medium · 15–29 = small · under 15 = no number

• Source diversity

  4+ source types and the largest ≤ 60% = diverse · 2+ types and the largest ≤ 85% = limited · otherwise = one type dominates

• Model-to-model tagging consistency

  From the consistency test (Cohen's kappa for content tone between the production model and a reviewing model on a sample — not accuracy proven against human labels): 0.6+ = good · 0.4–0.6 = moderate · under 0.4 = weak; one level lower when over 30% of the area's content is jokes/sarcasm (harder to classify)

• Score stability

  We recompute the area's number 1,000 times on random resamples of the same material: 90% range ≤ 6 points = stable · 6–12 = moderate · over 12 = volatile

• External agreement

  We compare the area with sources independent of our collection and tagging: the tone of Egyptian news in GDELT (is the area heavier or lighter than the other areas, for us and for them?) and Arabic-Wikipedia attention (did it rise or fall on the same days?). All the same direction = agree · mixed = partly agree · opposite = diverge · not enough outside signal = —. Agreement is not proof and disagreement is not an error: it tells you how cautiously to read the number

And the weekly number (51.4): recomputed on random resamples of the same material, 90% of results fall between 50.5 and 52.3. A one- or two-point change between weeks is not necessarily real. This range reflects uncertainty in our sample and calculation only — not in Egyptian public opinion.

Consistency test

We re-checked the tagging on a sample — with a second model

Important: we do not yet have an independent human benchmark. This is consistency between the production model and an automated reviewer — not proven accuracy. Next step: a human-labelled sample.

We drew a random sample from the week (spread across source types), had a second, stronger model (Claude Opus) tag it from scratch without seeing the original tags, and compared.

The two models agreed on the tone of 7 in 10; kappa 0.55 = moderate agreement, with no meaningful systematic lean.

Limits of this test

• The second annotator is another AI model (Claude Opus), not a panel of human Egyptian readers — agreement shows consistency between models, not proven correctness.
• Only one second annotator; models from the same family may share blind spots, so agreement can be inflated.
• a random sample from one week (2026-10-25 → 2026-11-01); small groups (e.g. Social posts, Videos / songs / trends) give wide error margins.
• Egyptian dialect, sarcasm and very short comments are hard for any reader; some items are genuinely ambiguous.
• The second annotator saw the same short text the production tagger saw (titles/snippets, not full articles, videos or audio).
• Not validated here: Egypt-relevance filtering, story grouping, source weighting, the index formula, lenses, joke types, or whether the collected items represent all Egyptians.

Full numbers for every area on the numbers page; principles on About.