← Playground一覧に戻る

Box Score

The same run of machine-written stories as the Morning Edition sheet, taken as figures instead of prose — every score distributed, averaged and ranked on one statistical page, where choosing a measure redraws the field and choosing a story follows it through every plate.

Box Score

Nos. 01–20 2026.08.14 → 2026.08.24 20 stories rubric v2.0

Figures for the stories on the Morning Edition sheet.

Measures

Order

  • Craft read by a critic, out of ten a side
  • Creativity counted from character bigrams, out of a hundred
  • Making the record of how the story was produced

Craft 合計 /70

n
20
mean
58.85
median
59.0
σ
1.62
range
56–61
σ/mean
2.8%

Highest

  • 61
  • 61
  • 61
  • 61

Lowest

  • 56
  • 56

Columns: mean, ±σ, and σ as a share of the mean — the last one being the only figure that can be compared across measures of different sizes. Of the numbers that judge a story, 想起性スコア moves most and 日本語の自然さ least — a measure at the bottom of that order has said the same thing about every morning.

Standings

Head a column to sort by it, or No. to go back to the mornings. The # column is the rank in whichever measure is open above. A title opens the story itself.

Story Craft Creativity Making
# Title
1 17 61 9 8 9 9 8 9 9 63 95 82 12 981 0
1 16 61 9 8 9 9 8 9 9 64 92 85 16 989 12 0
1 02 61 9 8 9 9 8 9 9 61 92 76 15 1120 11 0
1 01 61 9 8 9 9 8 9 9 67 94 84 22 1116 19 1
5 13 60 8 8 9 9 8 9 9 63 90 79 19 1036 11 0
5 11 60 9 8 8 9 8 9 9 66 94 84 19 1155 10 0
5 08 60 9 8 8 9 8 9 9 61 92 75 16 1028 5 0
5 05 60 9 8 8 9 8 9 9 66 94 88 16 1104 15 0
9 19 59 9 9 8 9 7 9 8 62 91 74 21 1090 10 0
9 14 59 8 8 9 9 8 9 8 59 92 75 11 957 9 0
9 12 59 8 8 9 9 8 9 8 70 99 98 14 1123 9 0
9 04 59 8 8 8 9 8 9 9 63 91 79 19 1103 12 0
13 20 58 8 8 8 9 8 9 8 59 92 73 12 831 9 0
13 10 58 8 8 8 9 8 9 8 62 93 76 16 1081 15 0
13 03 58 8 8 9 9 7 9 8 59 92 74 12 1117 12 0
16 09 57 8 8 8 9 8 8 8 65 94 85 16 1006 10 0
16 07 57 8 8 8 9 7 9 8 64 92 84 17 1115 12 0
16 06 57 8 8 8 9 8 8 8 59 91 69 16 1042 13 0
19 18 56 8 7 8 9 7 9 8 63 93 73 23 938 10 0
19 15 56 9 7 8 9 7 8 8 62 89 77 19 1107 12 0

合計 /70 物語 物語の完成度 /10 必然 展開の必然性 /10 文体 文体・文章の質 /10 日本語 日本語の自然さ /10 飛躍 独自性・飛躍 /10 継承 着想の継承 /10 余韻 余韻・主題の深度 /10 総合 総合スコア /100 非類似 非類似性スコア /100 独自 独自性スコア /100 想起 想起性スコア /100 字数 文字数 chars 時間 生成時間 min 改稿 改稿回数

Craft × Creativity

One judge read the story; the other counted its bigrams. r = +0.26

Creativity 総合スコア ↑
Craft 合計 →

The selected story

No. 20 Mon 2026.08.24

合計
58 13th
物語 物語の完成度
8 10th
必然 展開の必然性
8 2nd
文体 文体・文章の質
8 9th
日本語 日本語の自然さ
9 1st
飛躍 独自性・飛躍
8 1st
継承 着想の継承
9 1st
余韻 余韻・主題の深度
8 10th
総合 総合スコア
59 17th
非類似 非類似性スコア
92 9th
独自 独自性スコア
73 18th
想起 想起性スコア
12 17th

Rank out of 20. The tick on each rule is the field’s mean; the rule runs from the lowest figure in the edition to the highest. Read this morning →

No. 20 Mon 2026.08.24 831 chars Craft 58/70 Creativity 59/100

水をやる手

深夜二時、削除待ちの列にはまだ数百件が並んでいる。彼女はモニターの前で、規定に触れる語へ次々と下線を引いていく。〈死ね〉。〈お前の会社は潰れろ〉。〈これ買うと五キロ痩せました〉。〈○○さんの家はこのへんです〉。基準は明快で、迷う必要はほとんどない。手首の動きは机の木目のように均一になっていく。

棚の隅には、前任者が置いていった多肉植物の鉢がある。水やりの係は誰にも頼まれないまま彼女に回ってきた。去年の秋から花をつけたことはない。それでも枯れはしない。週に一度、コップの水を土にかける。それだけの関わりだった。

その夜、列の途中に一件、システムが判断を保留した投稿が挟まっていた。〈父は三年、便りをよこさない。便りを書ける場所にいないのだと、母は言わない〉。個人情報の項目には触れず、暴力的な語彙もない。ただ「収容」「拘束」といった語が離れた行に散っていて、自動判定が人の目に投げてきたのだった。

彼女はカーソルを止めた。学生の頃、詩のようなものを書いていた。書いたものは、実家の抽斗の底に何枚か眠っているはずだが、読み返したことはない。就職の面接で、その話をしたことは一度もない。

規定の欄には「政治的主張に該当するかは文脈を総合して判断する」とだけある。文脈を総合するための時間は与えられていない。次の投稿はもう列の先頭に並んでいる。カーソルは保留のボタンへ向かいかけ、そこで止まる。

指は、数百件を消してきたのと同じ角度で、いつもの位置へ吸い寄せられていく。理由欄には「関連語ヒット」とだけ入力した。手つきは、〈死ね〉を消した時と変わらなかった。投稿は列から消え、次の行が画面を占める。

作業を終えたのは明け方近くだった。彼女はコップに水を汲みに立ち、多肉植物の鉢の前で足を止めた。土は乾いていた。水をかける。土の色が変わり、光る。蕾はどこにも見当たらない。彼女はコップを流しに戻し、また椅子に座った。換気扇の音が、規則正しく回り続けている。

Critique

深夜、投稿を機械的に削除し続ける女性の一夜を描いた作品です。反復する手つきの描写に静かな緊張が積み重なり、多肉植物への水やりという習慣が終盤で新たな重みを帯びます。抑制の効いた文体が読後にじわりと効いてきます。

This morning on the Morning Edition sheet →

No. 20 · 水をやる手

Box Score

Every morning a program reads three works from 青空文庫 and writes one short story of its own. It is then critiqued against a seven-axis rubric by something with the same faculties that wrote it, and separately measured against its sources by counting character bigrams. Both verdicts are printed on the Morning Edition sheet, beside the story they belong to.

This is the back page of the same paper. Nothing here is prose: it is the same run of mornings taken purely as figures — every score distributed, averaged, ranked and set against every other score, the way a paper sets out a league at the end of a week. Reading the stories tells you what each morning produced. Reading them as a table tells you something different and slightly rude: how much any of these numbers actually varies, and therefore how much of it is a judgement rather than a habit.

The three bands

BandWhat it holds
MeasuresEvery figure the stories carry — the seven craft axes and their total, the four creativity figures, and the record of how each was made — each with its mean, its standard deviation, and the ratio between them
StandingsAll the stories, all the figures, one row per morning, sorted by whichever column you head
Craft × CreativityThe two verdicts plotted against each other, with the correlation between them, and the full profile of whichever story is selected

The measures keep the names the generator gave them, which are Japanese: this is a rubric for Japanese prose, and translating its axes would quietly rename what they measure. The heads in the standings are short forms of those names, opened in full in the legend beneath the table.

The three bands are wired together. Choosing a measure on the left redraws the distribution, re-ranks the standings and moves the highlighted column; choosing a story — from a cell in the histogram, a dot in the scatter, a row in the table, or a name in the leaders — follows that story through all three. ← → step through the mornings from anywhere on the sheet; ↑ ↓ walk the measures. Every figure is rendered at build time, so with scripting off the page is still a complete statistical abstract, sorted by craft total.

Numbers are only half of it, so a title opens the thing they were taken from: tap one — in the standings, in the leaders, or on the selected story — and that morning’s text comes up over the whole sheet, set as prose is set on paper, with the critic’s verdict under it. ← → read on through the run without closing; Esc puts the sheet back exactly where it was.

Reading the distribution

The histogram is built out of the stories themselves rather than out of bars: one cell is one morning, printed at its own number, stacked in the bin its figure falls into. Where a measure is so concentrated that twenty cells pile into one column, the cells lose their numbers and the plot becomes a plain dot column — which is itself the finding.

MarkMeaning
The mean, on the axis
Dashed ruleThe median, through the plot
BracketOne standard deviation either side of the mean; an open end means it runs past the axis
Bar above the plotFor a measure with a ceiling, how much of that scale the run actually used

That last bar is the one worth watching. A seven-axis rubric marks out of ten a side, but a critic who never leaves a two-point band is not using the scale, and the bar says so at a glance before any of the other numbers are read.

Mean, spread, and what a number has to earn

Every measure is printed with its mean, its σ, and σ as a share of the mean. The first two cannot be compared across measures — a standard deviation of 2 means one thing out of ten and another out of a hundred — so the third column is the only one that can be read down. Order the measures by spread and the list re-sorts by exactly that ratio, from the figure that discriminates most between mornings to the figure that discriminates least. A measure at the bottom of that order has, over the whole run, said the same thing about every story. It costs as much to produce as the one at the top.

The standings then let you check the same thing by hand: head a column and watch how little the order changes.

Craft × Creativity

The two verdicts are plotted against one another, craft along the bottom and creativity up the side, with a dashed rule through each mean and the Pearson correlation printed beside the caption. They are computed by processes that cannot see each other — one is a reading, the other is arithmetic over character bigrams — so there is no reason for them to agree, and the interesting stories are the ones far off the diagonal: well made and close to its sources, or far from its sources and poorly made. Selecting a point brings its full profile up beside the plot, each figure set on a rule that runs from the lowest in the edition to the highest, with a tick where the field’s mean falls.

Where the numbers come from

The same committed JSON file that sets the Morning Edition sheet. Nothing is fetched at runtime and nothing is computed in the browser except the standard score in the reading line: the means, the deviations, the bins, the ranks and the correlation are all worked out at build time, and the page ships as a printed table that happens to be able to re-sort itself. Re-export the stories with any number of mornings and this sheet takes as many as it is given — the bin widths, the axes and the ranking all follow the file.