The word well — a persona's own secrets

The house keeps one generic word bank, and any game that draws a secret draws from it: 西瓜, 雨伞, 地铁. It is correct, it is safe, and it is nobody's. The word well is the same machinery with a different author — every persona gets its own hundred words and its own hundred pairs, per language, so a secret drawn in their room lands as of course that's theirs. This page records the design, the laws the build paid for, the first qualification run, and the verdict. Built 2026-08-05 as groundwork for the 20Q / 卧底 overhaul; nothing is wired into a game yet — that is deliberately a later session's job.

1 · What the well is, and what it is not

lib/wordpools.py already solved the hard security half of drawing a secret: the room draws with a CSPRNG at apply time, after the speech has published, so the transcript never holds the word and therefore cannot leak it. The avoid-list, the placeholder resolution, the two pool names — all of that is finished and the well does not touch any of it.

What the well replaces is the authorship of the bank. One hand-curated general list becomes one generated list per persona per language.

The generic bank keeps a second job in the new arrangement, and it is the more interesting one: it is the negative control. A sample of it is pasted into every generator call and every judge call with one instruction — this is what a list looks like with no character in it; anything at this register fails. The bank is also checked mechanically, so an item that literally appears in it is dropped before it ever reaches the judge.

2 · The house grammar — the mind proposes, the world salts and picks

The dice specimen settled how this house builds anything that mixes code and a model: the world is code and the voice is a mind, and the seam is engineered so neither ever does the other's job. The well is that same split applied to authorship.

Why the salt exists at all

Ask a model twice for a hundred things from this character's world and you get nearly the same list twice, because a modal model has favourites and a persona profile does not dislodge them. So the server rolls the constraint before the call, from wheels the model cannot see the shape of:

• facet — what kind of thing (28 of them: something held in one hand · something that gets thrown away · something nobody is supposed to touch). Every facet is deliberately generic: a wheel with "something from a kitchen" on it would hand the chef its answer and starve the designer.
• scope — which slice of a life to look in (10: from their working hours · from what they argue about · from their failures and the parts they would rather hide).
• sound — a soft quota on the initial (three pinyin initials for Chinese, three letters for English). It is a nudge, not a cage; a miss costs nothing and a hit knocks the model off the first word it reaches for.

Both wheels are drawn without replacement and reshuffled when spent, so a full well walks through most of the facet space instead of circling one corner of it.

3 · Two kinds, and why the pair needs two minds

word is one secret for twenty questions: a countable, pointable thing, guessable by a stranger, tiered easy / medium / hard.

pair is two confusables for the hidden-role game — everyone gets word A, one player gets word B, everybody describes their own word, the table hunts the odd one out. A pair cannot be written by the persona call, and the reason is worth stating plainly: knowing a world and knowing what makes two words confusable are different jobs. The persona knows that a Shanghai property advisor lives among 板块 and 学区房; it has no particular skill at judging whether two words are equally describable under pressure. So the pair runs in two stages, with the persona out of the room for the second.

4 · The judge, and why the code holds the gate

A separate call scores every candidate on 0–3 axes. The judge scores; it does not decide. The keep is a threshold in Python, because a threshold is a dial we can turn after the owner eyeballs a sample, and a model's own verdict is a thing that drifts. The judge is still asked for a verdict — purely so the disagreement between its verdict and the gate can be counted.

Three tests the judge is given in words

• The table test decides concrete: if the table cannot ask is it bigger than a shoe? and get a straight answer, it is not an item. That kills every quality of a thing — 木纹 (wood grain), 粗糙边缘 (a rough edge), 哑光 (matte) — and every fitting that only exists as part of a bigger machine: a knob, a switch, a hinge. The table cannot bound a part.
• The proposer's reason is evidence, and the judge must check it. The generator returns a one-line why beside every item. If that line leans on would, probably, surely, must have — or paints a plausible life instead of citing a documented one — fit is capped at 1. This is the axis that catches a charming invention: "a small charcoal brazier would be a heat source used outdoors during tea" is a picture of a life, not a fact of one.
• Easy and fresh are not opposites. The first calibration produced wells with zero easy items, because the judge had quietly fused "a generic host would never write it" with "hard to guess". An oyster is easy to corner in twenty questions and unmistakably Bourdain's. fresh is about the author; tier is about the guessing.

5 · The free half — what never reaches the judge

Four mechanical filters run before any scoring call. They cost nothing, they never drift, and between them they catch the three classes of defect the judge was measurably soft on.

6 · The no-padding law

It is enforced in four places, and the fourth is the one that took a rewrite:

1. The generator is told, in as many words, that a short honest batch is the correct answer and that an item a generic host would also have written is a failure rather than a filler.
2. The judge's fresh axis is anchored on the house bank as an explicit negative control, sampled fresh into every call so the model cannot learn to dodge a fixed list.
3. The loop stops after four consecutive barren salts and reports capped at N.
4. A barren salt is not a persona's ceiling. "Something people wait for or queue for, from a designer's failures" comes back empty for almost anybody — that is a fact about the wheel. The first calibration treated any batch yielding fewer than two items as barren and gave up after three, which capped Dieter Rams at eight. Barren now means contributed nothing at all, and the verdict needs four different salts in a row to agree.

The audit therefore separates why a run stopped: full · dry (the ceiling — four barren salts) · batch_ceiling (the budget ran out first, which is a fact about the run, not about the persona). Conflating the last two would let a cheap run libel a rich persona.

7 · The qualification run — 2026-08-05

Three personas, chosen to be as different as the roster allows, each asked for 100 words and 100 pairs in both Chinese and English — twelve wells from one command.

• Anthony Bourdain — chef and travel host. A huge, sensory, heavily documented life; the easy case.
• Dieter Rams — industrial designer. A narrow world of objects and refusals, and every one of those objects is famous, which turns out to matter.
• 小胖 (Xiao Pang) — a Shanghai property advisor built from a supplied transcript corpus. The deliberately hardest case: no public biography, and a world made of advice rather than things.

Cost and shape of the run: 1,406 API calls on deepseek-v4-pro, $1.09 for all twelve wells, ~25 minutes of wall clock at twelve-way concurrency. 4,462 candidates were proposed; 645 were dropped as duplicates or near-variants, 169 for unplayable shape, 44 for being literally in the house bank, 521 had fit capped by the hedge scan, and the judge cut 2,968.

The second pass — the same wells with a wider stopping rule

Bourdain returning 97 words in Chinese and 28 in English, from one profile under one rubric, is not a fact about his life. So the six word wells were re-run with --dry-streak 10 instead of 4 — one dial, nothing else changed:

The score distributions

Coral bars are at or above the gate. The mass sits below it in every case, which is what a working judge looks like — and the shape of the tail is the interesting part: it is long and thin for words (a few outstanding items, a wide field of near-misses) and fatter for pairs, where a candidate that survives the seam and symmetry tests tends to survive everything.

Anthony Bourdain — Chinese: 97 words · 76 pairs

Anthony Bourdain — English: 28 words · 84 pairs

Dieter Rams — Chinese: 26 words · 51 pairs

Dieter Rams — English: 13 words · 49 pairs

小胖 — Chinese: 7 words · 49 pairs

小胖 — English: 10 words · 27 pairs

8 · What the run found

Finding 1 · The first measurement was measuring its own stopping rule

This is the finding that reorganised every other one, and it arrived by accident: the 97-vs-28 gap in Bourdain's two word wells was too large to be a fact about a person, so the stopping rule went on trial. It lost. Raise --dry-streak from 4 to 10 and Rams' English well goes 13 → 49 and Bourdain's 28 → 80, both still climbing when the batch budget cut them off, both at a lower cut rate than before. Pass 2 cost $0.49 and was worth more than the $1.09 that preceded it.

The consequence for everything else on this page: every "capped at N" is a floor, and a soft one. The quality of what survived is measured; no ceiling here is. Two things were already built for this and both earned their keep — stop_reason separates a dry stop from a budget stop, and --dry-streak is now the dial. Neither is calibrated yet.

Finding 2 · The pair well looked twice as deep as the word well. It was an artifact — but the design observation underneath survives

Pass 1 read: words averaged 30, pairs averaged 56, and the obvious story was that pairs are structurally easier to fill. That comparison was rigged and I did not notice — every word well was stopped by the streak while every pair well ran the full 90-batch budget. At equal budget the gap closes to nothing:

Rams · English — 49 words, 49 pairs. Bourdain · English — 80 words, 84 pairs.

What survives is not a number but a mechanism, and it is still worth carrying into the overhaul. The two judges ask different questions about freshness:

the word judge asks would a persona-less host have written this WORD? · the pair judge asks would a persona-less host have built a pair in this AISLE?

收音机 alone fails — every generic bank has "radio". 收音机/录音机 passes, because a man who spent a career designing radios owns that aisle and the second half is supposed to be ordinary. The pair form lets the persona supply the context while the words stay common, and common words are what a party game needs. An oyster in Bourdain's room is a good secret even though anyone might have picked "oyster" — what makes it his is that the questioning will be about him. Whether re-anchoring word freshness on the aisle actually helps is now an open experiment rather than a conclusion, because the evidence that suggested it turned out to be the stopping rule.

Finding 3 · The personas do hit different walls — but read the rate, never the total

Absolute survivor counts are contaminated twice over — by Finding 1 and by plain sampling noise. The share of proposals killed on each axis is a rate over hundreds of candidates, so it is the one thing worth reading. The table below is pass-2 data (700–800 proposals per well rather than pass 1's 84–236), because a finding built on the thinner half of the evidence is how this page went wrong the first time.

Note how much the small-sample rates moved: 小胖's not-a-thing rate read 31% on 86 pass-1 proposals and 17% on 236 pass-2 ones. A rate is only stable once the denominator is.

What is left, after both corrections, is not a rate at all — it is persistence. Bourdain and Rams ran 90 batches still producing, in both languages. 小胖 ran 27 and 28 and stopped. His word well is the only one in the study that has been shown to have a bottom, and meanwhile his pair well reached 49 and was still budget-limited when it stopped. That within-persona gap — 14 words against 49 pairs — is the real evidence for the seating rule, and it is the opposite in kind to the across-the-board claim withdrawn in Finding 2.

Finding 4 · The judge's rejects are the evidence that it works

Cut rates of 41–80% are only defensible if the cuts are right, so the wells ship with them (--report --cuts N). They read correctly. From Rams' English run:

The judge disagreed with the code's gate 227 times across the run and the code won every one of them, which is the whole reason the threshold lives in Python.

Finding 5 · Three defects the run exposed and this build has not yet fixed

1. Invented and transliterated words survive. Bourdain's Chinese well contains 奇布 (a mangling of 旗袍), 福轿 and 莱斯霍 — a transliteration of Les Halles, which is a proper name wearing Chinese characters. All three scored 9/9. The shape filter cannot see them (no digits, no latin, short enough) and the judge did not apply its own proper-name rule. Xiao Pang's 静安府 is the same defect: a housing complex is a place with a name, not a secret. The fix belongs in the judge's concrete anchor — if the word is a transliteration of a foreign name, or a coinage no dictionary holds, it is not a word players share.
2. The modifier hack survives at short lengths. The rule caught 日式盆栽剪刀/日式裁缝剪刀 (the shape filter kills anything that long), but 铜铰链/铁铰链 scored 13/15 and 磨刀石/磨刀棒 11/15. Both are one noun with a material swapped in front. The hard part is that 收音机/录音机 has the same surface shape and is a good pair — the difference is that each half of that one is a word in its own right. The examples in the judge's prompt are all long; they need short ones.
3. A place can fail the table test by mistake. Rams' garden was cut with "a place, not a thing" at fit 3 · concrete 0 — but the rubric explicitly admits places you can walk into, and 图书馆 and 动物园 are sitting in the house bank right now. One clause in the concrete anchor needs to be louder.

All three are fixed as of the commit that carries this page, and none of them was fixed in the build that produced the numbers above — changing the rubric mid-measurement would have voided it. The judge's concrete anchor now carries an explicit a place is a thing clause and an it must be a word players already share clause; the modifier-hack rule now names short examples and states the real test (is each half a word in its own right? — 收音机/录音机 yes, 铜铰链/铁铰链 no); the generator is told never to transliterate or coin. None of the three fixes is measured — that is the next run's job.

9 · The verdict

The dials, in order of leverage

1. Calibrate the stopping rule first, before believing any other number. Everything downstream — how deep a persona goes, whether words or pairs fill faster, whether an advisory seat is unfit for 20Q — was distorted by it once already. Replace the streak with a plateau test (stop when the last K batches added under X% of target) so the rule scales with yield instead of sampling it.
2. Feed the well the corpus/. It currently reads profile.md + tags.txt and nothing else — about 30k characters. Bourdain has a whole corpus directory of his own sentences sitting unused, and not-theirs did a quarter to a half of all the cutting; 小胖's English well cut 54% of proposals that way. This is the largest untapped evidence source and it costs only prompt budget, which prefix caching has already made cheap.
3. Then re-ask the freshness question for words — anchored on the aisle rather than the bare word, the way the pair judge already does. Worth running as a clean A/B now that the evidence which suggested it has been withdrawn.
4. Let the target be per persona. The machine already ships what survives; the change is in how a short well is read — provided the stopping rule has been calibrated first, or a short well means nothing at all.
5. Seat by well, not by taste. Generate before choosing a game for a persona — and read stop_reason, not the count. A seat that goes dry under a patient stopping rule has told you something true about its world; a seat that hits batch_ceiling has told you only that you stopped paying.

What the 20Q overhaul inherits

1. A number a stopping rule produced is a number about the stopping rule until you have re-run it with that rule moved. Twice on this page a confident finding was really an artifact of when the loop gave up — including the one this page originally called its most useful result. Any generate-until-exhausted loop needs its termination condition varied before its output means anything.
2. The draw does not change. wordpools.py keeps the CSPRNG, the apply-time draw and the avoid-list; a well is a bigger, better-authored pool behind the same door.
3. Every well carries its own audit — salts used, cut rate, blocking axis, stop reason, cost, and the rejects. The rejects are what let the rubric be argued with, and the blocking axis is what caught two wrong diagnoses; without them the only way to tune the gate is to guess.
4. The persona proposes; the world salts and picks. Where a model has favourites, roll the constraint in code before the call — and roll one constraint, because a conjunction of three starves the answer.
5. A thing a game can use is not the same as a thing a persona owns. The table test, the shape filter and the hedge scan all exist because the first drafts confused the two, and every one of them is cheaper in code than in a prompt.
6. Wire nothing yet. The wells in rooms-dev/_wordwell/ are a measurement taken with a mis-set dial and three since-fixed defects — not a shipping artefact.