The word pool — one pool, every persona

The word well generated a bespoke set of secrets for each persona: $0.50 and twenty minutes each. That cannot survive a roster of user-made characters — nobody waits twenty minutes to play with someone they invented thirty seconds ago. So the authorship moves off the persona entirely. One pool is built once for everybody; a persona is served by ranking that pool against its own card. Per persona: one embedding call and a dot product.

1 · The shape

2 · What the build does, and what it costs

3 · The curation pass — the only reason this works

Concreteness ratings cannot make a pool playable on their own, and the reason is sharp: the worst possible secrets are perfectly concrete.

Three classes die in the curation pass, and all three were found by looking at real ranked output rather than by guessing:

• people and roles — chef · realtor · engineer · 厨师 · 顾问 · 检查员
• abstractions and screen-words — world · bottom · item · goods · 世界 · 底部 · 物品 · 帖子 · 页面 · 视频. A party game is about the physical world.
• proper names and breeds, which the source lexicons carry as ordinary entries — 知味观 · 谭鱼头 · 欧宝 · 解百纳, and the sneakiest of all, breed names: 浦东鸡 · 小萨摩 · 珍珠狗 · 维兹拉犬. "chicken" keeps; "浦东鸡" drops. These survived a first curation pass that did not name them, which is why the prompt names them now.

The pass keeps about 63% and runs in under a minute for the whole pool, once.

4 · Chinese has no foundation — so translate, don't scrape

English has a genuine free foundation. Chinese does not, and this was the sharpest finding of the whole exercise.

Ranking a designer against a pool of animals, food and cars produced exactly what you would expect:

苹果 欧宝 加去 新灰 欧姆贝 莱姆 德国黄牛 颐达 灰插 嘉好 全保 小灯 力存 黑漂
— Dieter Rams, ranked against THUOCL. Car models and noise.

Translating the concreteness-validated English pool fixes it, and the Chinese side inherits Brysbaert's validation for free:

家具 设备 椅子 机器 工厂 桌子 工具 书桌 沙发 皮革 杯子 橱柜 铅笔 床 木头 衣柜
— the same persona, the same query, after translation.

Attribution settles the recipe. Of the junk left in the top thirty, eight of twelve came from the THUOCL scrape and none from the translated layer. Only the very top of the food and animal lists is kept, for the words translation can never reach — 火锅, 豆腐, 包子 do not fall out of an English pool.

5 · The query is a world, not a name

6 · What it picks

Three personas chosen to be as different as the roster allows, each ranked against the shipped pool. Nothing was generated for them; every word below was already in the pool before any of these characters were considered.

Anthony Bourdain — chef & travel host

Dieter Rams — industrial designer

小胖 (Xiao Pang) — Shanghai property advisor

7 · The owner's sweep — "not too niche, and not a broad category"

The owner eyeballed the personas starting with A and deleted words on one rationale, stated as two rules that are really one rule seen from both ends:

they have to be common ideas or objects, not too niche — and they can't be a very broad category such as 工具, humans are impossible to guess

The sweep, and the mistake it took to get right

The first attempt cut 16–20% and was wrong: it dropped 厨房 · 考拉 · 猎鹰 · 瓢虫 · mountain · dragon · paper. It had confused broad with not a small handheld object. The corrected test asks the only question that matters:

Could two players, both correct, be picturing completely different objects? "工具" — one pictures a hammer, another a screwdriver: drop. "考拉" — everyone pictures a koala: keep. It is a question about whether the word SPANS DIFFERENT OBJECTS. Not about size, and not about whether it fits in a hand.

With that, the sweep takes 4–5% of words and 10–14% of pairs — categories (配饰 · 装饰品 · 餐具 · appliance · attire · condiment), genuine niche (铃舌 · 雾号 · plinth · mollusk · plumage), and pairs that are merely synonyms (house/home · 草地/草坪 · port/harbor · 腊肠/香肠), which have no seam and so can never end a round.

Then the same rule again, measured across the whole roster

Per-word judgement cannot see the version of "too broad" that actually hurts. Rank the pool for all 105 personas and count how often each word lands in a top-30:

家园 ×75 · 图片 ×71 · 雪球 ×43 · 绿洲 ×36  |  globe ×59 · top ×59 · soapbox ×54 · nutshell ×40
— out of 105 personas, including a chef, a designer and a mathematician.

A word in the top-30 of a third of every character in the product is about none of them. Some are class words; most are semantic attractors — words near the centre of the embedding space, or ones a biography reaches for metaphorically (soapbox, nutshell, 绿洲, 家园) — and they crowd out what actually belongs.

What that buys, persona by persona — 小胖 keeps 房子 at number one and gains 别墅 · 公寓 · 顶层公寓 · 联排别墅 · 平房 · 阁楼 · 屋顶; Rams goes from 装饰品 · 图片 · 基座 to 高脚椅 · 削皮器 · 咖啡壶 · 洗碗盆 · 扶手椅 · 床头柜 · 脚凳. The penalty is one number (MAD_WORDPICK_BREADTH, default 0.7); above ~0.9 it starts pulling in noise.

8 · The trade, stated plainly

Overlap between what this picks and what the word well generated is 27–36%. Two thirds of the generated words are unreachable from any generic pool, and they are the good ones: 槽钢层 · 半阳台 · 蛏刀 · 透明盖 came out of a persona's own documents and no shared list will ever hold them.

The word well is not deleted. It stays as the deepening path for a persona somebody cares enough about to wait for, and its findings — the stopping-rule trap, the seating rule, the pair rubric — are what made this build possible in an afternoon.

9 · What is not done

1. Nothing is wired into a game. This ships the pool and the picker; the 20Q and 卧底 overhaul is the next session's job, and it should read §6 and §7 before choosing a target size.
2. The picker is unmeasured on real play. The samples below are eyeballed, not played. The word well's own lesson applies: the sims cannot find what only happens to real people taking real time.
3. Relevance is topical, not characterful. Ranking cannot distinguish a word the persona owns from a word merely in their trade. If that turns out to matter at the table, the fix is a cheap re-rank of the top ~200 by one model call — still once per persona, still no build step.
4. Licences want checking before this ships to production. THUOCL grants commercial use explicitly; the Brysbaert mirror states none; SUBTLEX-CH says research only and is not used here. The translated layer is our own output.