Files
wikideck/docs/game/CARD_DATA.md
T

8.5 KiB
Raw Blame History

Card data: real stat distributions (council-data)

Status: interim, generated 2026-11-16 23:06 EST. The full 330-card run is still fetching pageviews politely; this file and card-sample.json will be regenerated.

Method: scripts/night/card_sample.py draws random ns0 articles in batches of 20 and keeps those passing the server/wiki.ts filter (thumbnail from upload/thumb.wikimedia.org, intro extract of at least 40 characters). It enriches them with the same API parameters as server/enrichment.ts: six completed months Jun–Oct 2026; attention uses October. It adds Wikidata P31 and 30-day drift (views 10-17..11-15 vs 09-17..10-16). scripts/night/card_analyze.py produces the tables below.

Headline findings (INTERIM: 100 fully enriched random illustrated cards; 330 sampled, rest still fetching under Wikimedia 429 rate limits)

  1. Raw stats are compressed and skewed. Median card: attention 33, depth 63, reach 24 (p10–p90: attention 23–53, depth 54–71, reach 9–48). No card above 81 on anything; 100 is unreachable honestly. Use the 1–10 tier tables below (sample deciles), not 0–100.
  2. Correlation is weaker than feared, except attention↔depth (Spearman 0.59). Reach is nearly independent of both (0.10 / 0.02): stub villages get many bot-made language editions. So reach is a real separate axis; attention+depth partly double-count fame.
  3. Underlying raw values are very long-tailed: monthly views median 94 (p90 1,438, max 14k); pageBytes median 6 KB; languageCount median 4, and 10% have 0 other languages (reach 0).
  4. Trend is mostly negative: 69% of cards have trendPercent < 0 (median −10%), 31% positive, 14% swing by more than 50%. Null is rare (1%). A "Trending" keyword fires for about a third of cards: fine as a small ±1.
  5. Live values are fairly stable in score terms: 30-day pageview drift median |18%| (p90 55%), but that moves the attention score by only a median of 1.4 points (p90 5.6). So "live vs. mint" differences are small for random cards; big swings only come from deliberate inflation (see the critic's point A4).
  6. The legacy category() buckets are unusable as factions: Culture 53%, People 30%, Nature 9%, History 5%, Science 3%, Space 0%.
  7. Candidate factions from Wikidata P31 (instance of), clean and readable: People 31%, Places 25%, Nature 10%, Structures 6%, Works 5%, Events 5%, Sport 3%, Ideas & Tech 3%, Organisations 2%, unclassified 10%. Suggest 6 factions: People, Places, Nature, Works & Culture (Works + Organisations + Sport), Structures, Events & Ideas (Events + Ideas & Tech), plus a neutral fallback. People and Places are big, so faction bonuses should scale with share in deck, not raw count, or the small factions need their own perks.
  8. Starter/loaner list: frozen stats for up to 4 cards per legacy category, spread across tiers, in docs/game/starter-cards.json (Space has no sample cards: hand-pick e.g. the Moon from shared/catalog.ts).

Full tables

Sample: 100 cards; ~44% (149/340 logged during sampling) of random pages pass the illustrated filter. Pageviews available for 100.

stat n min p10 p25 median p75 p90 max mean skew
attention 100 13 23 27 33 41 53 69 35 0.75
depth 100 50 54 59 63 68 71 81 63 0.24
reach 100 0 9 16 24 36 48 59 26 0.11
pageBytes 100 998 1804 3407 6018 12512 19161 74169 9885 3.20
languageCount 100 0 1 2 4 11 27 56 9 1.83
monthlyViews 100 5 22 43 94 277 1438 14374 611 5.76
dailyAverage 100 0.2 0.7 1.4 3.0 8.9 46.4 463.7 19.7 5.76
trendPercent 99 -86.6 -45.1 -26.4 -10.4 4.5 21.1 187.1 -9.5 1.83
drift30Percent 99 -94.7 -45.9 -30.7 -12.7 -1.7 14.8 72.1 -16.1 -0.20

Spearman rank correlations (cards with pageviews):

attention depth reach trendPercent drift30Percent
attention 1.00 0.59 0.10 0.24 0.25
depth 0.59 1.00 0.02 0.07 0.05
reach 0.10 0.02 1.00 -0.10 0.24
trendPercent 0.24 0.07 -0.10 1.00 0.53
drift30Percent 0.25 0.05 0.24 0.53 1.00

(n=99; attentionmonthlyViews, depthpageBytes, reach~languageCount are monotonic so their rank correlations are identical.)

trendPercent sign: positive 31 (31%), negative 68 (69%), zero 0; null (prior month 0 or no views) 1. |trend| > 50%: 14; > 100%: 1.

30-day drift (views 10-17..11-15 vs 09-17..10-16): median |drift| 18%, p90 |drift| 55%. Attention score change implied: median 1.4 points, p90 5.6 points.

attention histogram (deciles 0-9,10-19..90-100): 0 4 30 39 12 11 4 0 0 0 depth histogram (deciles 0-9,10-19..90-100): 0 0 0 0 0 30 53 15 2 0 reach histogram (deciles 0-9,10-19..90-100): 10 25 23 23 10 9 0 0 0 0

Reach = 0 (no other languages): 10; reach 100 (>=999 langs) impossible in practice, max langs 56. Attention null: 0; attention 0: 0.

Sum of three stats: min 84, p10 97, p25 107, median 120, p75 140, p90 155, max 193

Legacy category() (server/wiki.ts): Culture 53, People 30, Nature 9, History 5, Science 3

Top Wikidata P31 classes: human 30, taxon 8, river 4, village 4, commune of France 3, musical group 2, building 2, castle 2, rural municipality of Austria 1, municipality of Austria 1, municipality of Brazil 1, television film 1, nor'easter 1, rail company (Japan) 1, village of Poland 1, municipality 1, civil parish 1, association football match 1, final 1, local government area of Victoria 1, plant disease 1, basketball team season 1, mosque 1, functional programming language 1, purely functional programming language 1, dependently typed programming language 1, programming language 1, free software 1, proof assistant 1, municipality of Croatia 1, municipality of Bangladesh 1, town 1, town in the United States 1, municipality of Georgia 1, epiphyte 1, protest 1, region of Mauritania 1, legislative election 1, musical work/composition 1, plateau 1

No Wikidata item: 0; no P31: 6

Most frequent category words (cards containing): stubs 35, births 27, people 26, from 24, century 22, living 16, geography 15, deaths 14, american 13, county 9, university 8, male 7, united 7, film 7, players 7, district 6, states 6, sports 6, biography 6, television 5, alumni 5, genera 5, established 5, villages 5, football 5, city 5, buildings 5, expatriate 5, sportspeople 5, sportsmen 5, towns 4, municipalities 4, writers 4, women 4, faculty 4, communes 4, rivers 4, places 4, establishments 4, actors 4, california 4, russian 4, association 4, structures 4, described 3, british 3, french 3, ohio 3, academics 3, based 3, japanese 3, pennsylvania 3, argentine 3, australia 3, populated 3, directors 3, rock 3, musical 3, conference 3, division 3

Rank-normalised tiers (1–10)

Tier = 1 + number of breakpoints the score is >=. Breakpoints are the sample deciles p10..p90 of the 0–100 score. Ties make some tiers thin (scores are integers).

stat p10 p20 p30 p40 p50 p60 p70 p80 p90
attention 23 25 28 31 33 36 38 45 53
depth 54 57 60 61 63 65 67 69 72
reach 10 10 16 20 26 30 36 39 49

Resulting tier counts — attention: None:0 1:10 2:9 3:7 4:14 5:8 6:12 7:7 8:12 9:10 10:11 ; depth: 1:7 2:11 3:12 4:6 5:11 6:11 7:9 8:9 9:14 10:10 ; reach: 1:10 2:0 3:13 4:12 5:15 6:8 7:11 8:10 9:11 10:10

Starter candidates: 19 written to /tmp/claude-1000/-home-eric-worker-jobs-night-council-data/d4e31987-ce7f-4f28-8262-8f0225215102/scratchpad/starter.json: Counter({'Nature': 4, 'People': 4, 'History': 4, 'Culture': 4, 'Science': 3})

Candidate factions (Wikidata P31 rules)

faction n share median attention median depth median reach example
People 31 31% 33 65 26 Salahuddin of Aceh
Places 25 25% 28 61 36 St. Georgen am Längsee
Nature 10 10% 30 56 31 Dichromorpha viridis
Other 10 10% 52 64 22 Village (Pennsylvania)
Structures 6 6% 32 58 16 Pathrail Mosque
Works 5 5% 40 64 16 The Best of Men
Events 5 5% 38 72 10 April 2021 nor'easter
Sport 3 3% 25 71 0 1912 Argentine Primera División final
Ideas & Tech 3 3% 37 62 20 Agda (programming language)
Organisations 2 2% 50 62 24 Hiroshima Electric Railway

Unclassified P31 labels: typeface 1, annexation 1, position 1, family 1