The word you were going to say
A machine seals three guesses at the first word you will think of, then grades itself. On the same 861 real first responses it scores 74.7% when it is allowed to pick its three words from the answers it is about to be marked on, 63.9% when one answer is held back, and 55.2% when the three words come from a different country and a different decade. Nothing about the machine changes between those three numbers. Only how much of the answer it saw first.
Nothing you type leaves your browser. There is no server, no analytics and no request to any other host: the whole dataset is inside this page, and you can read the source to check that.
It commits first, in public.
The box below is dead until a SHA-256 digest of the machine's guesses is on screen. Afterwards it shows you the full text that was hashed and re-checks the digest in front of you, and gives you a one-line shell command so you can check it somewhere this page has no vote.
Disabled until the digest below exists. One word, or a short phrase. Punctuation and capitals are ignored.
Waiting for the seal.
SHA-256 over the machine's guesses, the dataset digest, the matching rule and the scoring rule. Shown before you may type.
computing
Ten rounds is almost no evidence. 7 of 10 has an exact 95% interval of 34.8% to 93.3%, which overlaps everything this page could possibly claim. The interval is printed with every rate for that reason, and every miss is kept.
Same machine. Three scorings. Watch it fall.
Sixty native English speakers at Lancaster University were given category names and 60 seconds each to say as many members as they could. Each category was seen by twenty of them, and Banks and Connell recorded which member each participant said first. That is 861 first responses across the 44 categories used here, and it is the only thing being predicted anywhere on this page.
Pick who is allowed to choose the three words. Everything else is held fixed: the same 44 categories, the same 861 responses, the same lemmatisation, the same tie-break, the same denominator.
Why the big number is the dishonest one
In-sample, the three words are selected using the very observations that then grade them. With twenty people per category the tail is nearly empty, so a word said first by exactly one person can be pulled into the top three and score a guaranteed hit on that one person. -- of these 44 categories reach a perfect in-sample score that way, and -- of all 117 do. The clearest case is not in this comparison set at all: for four-wheeled vehicle, the in-sample top three is --, where the second and third words were each said first by exactly one person out of twenty. A top three that includes benz is not a fact about people. It is a fact about an empty tail.
Leave-one-out removes that specific cheat and costs -- percentage points. It does not remove the rest: the nineteen people who still choose the words were in the same room, in the same month, in the same university town. Going fully out-of-sample removes that too, and costs another --.
What each number is a number about
- In-sample is not a prediction at all. It is a description of twenty people, phrased as though it were about you.
- Leave-one-out is a real prediction about a twenty-first Lancaster undergraduate in 2017.
- Fully out-of-sample is a real prediction about a person in a different country, thirteen years later, from words chosen by hundreds of Americans who never met them.
- None of the three is a prediction about you. You are one draw, from a population nobody sampled, doing a task nobody ran. The page says so again below, at more length, because it is the honest limit of the whole exercise.
Denominator: every first response recorded in these 44 categories, -- of them. -- of the 44 categories have fewer than twenty, because a participant was excluded or gave no scoreable first word, so counting a flat 20 per category would invent -- misses. Dropping the -- first responses that nobody else in the room repeated would delete that many guaranteed misses and flatter every rate by about three points. All three denominators appear in the grid below, so you can see what each is worth.
Four free choices, and what each is worth.
Every number above depends on four decisions that a reader has no way to see from the outside. Rather than assert that ours are the right ones, here is the whole grid, recomputed live. The ordering never changes: in-sample is always the largest and fully out-of-sample always the smallest, under every combination. The level moves by up to five points, which is exactly why a single unqualified percentage is not an honest thing to print.
| Word merging | Categories | Denominator | Tie-break | In-sample | Leave-one-out | Out-of-sample | n |
|---|
The one that manufactures hits
Banks and Connell published their norms twice: a referential version that groups apple with apples, and a full version that keeps every spoken form apart. Switching between them, changing nothing else, moves the out-of-sample rate by -- points. That is a linguistic judgement made by two researchers in 2018, and it is worth more than most of what a predictor could do.
Which is why the matching rule is inside the seal, not applied afterwards. A commitment to an answer, without a commitment to how the answer will be judged, is not a commitment. If the rule can be chosen after seeing what you typed, the reveal cannot lose.
The rule is deliberately small. It folds apples onto apple and nothing else. It does not know that mum is mother, or that an American sailboat is a British sailing boat, and those failures are visible in the scores rather than patched away. Across the 44 categories the fold merges exactly -- pairs of distinct listed members, and every one of them is a singular and its own plural.
The one nobody mentions
Ties. With twenty people per category, third place is often a tie between two or three words that each got one first mention, and something has to break it. This page breaks ties alphabetically, which is arbitrary but is the only rule available that consults no further data: a tie-break by production frequency would, under leave-one-out, quietly look at the very observation being held out.
Breaking ties the other way instead moves the out-of-sample count from -- to -- hits out of 861, worth -- points. -- of the 44 categories have a tie straddling third place.
What a different scout got, and why
This page was specified from an earlier survey that reported 74.7%, 60.9% and 54.1%. Two of those three came from a different row of the grid above: the 60.9% is the verbatim word merging over all 117 categories, while the 74.7% is the referential merging over 117 categories with a flat denominator of 20 per category. Read down a single row instead and the fall is --. The story survives the correction. It was just being told with three different instruments.
Before it is pointed at you, it reproduces a printed result.
Banks and Connell compared their own norms against two earlier sets and printed six correlation coefficients. The same unmodified code that scores your word recomputes all six here, live, from the comparison tables they published alongside the paper, using Spearman's rho with average ranks for ties and pairwise-complete rows, which is what their own R script asks for. If any of the six disagrees with the printed value the row goes red and this page has a bug.
| Comparison | Measure | Printed | Recomputed | N printed | N here | Verdict |
|---|
The printed values are from the paper's Cross-region and Cross-time comparison paragraphs. Rounded to two decimals as printed; the recomputed column shows four so you can see the rounding rather than take it on trust.
Give it an opponent with nothing in it, and it should die.
A predictor that beats a cryptographic random number generator has a bug, not a talent. So the identical object that judges your word is set against two opponents that carry no information about human order, and it should collapse to the arithmetic baseline both times. If it does not, the scoring code is leaking the answer and this page has found a fault in itself.
A draw from the category's own vocabulary
For each category, crypto.getRandomValues draws one word uniformly from every member anyone produced for it, and that word is handed to the same judge() the game uses. The machine's guesses are unchanged and nothing is reset between rounds. It should hit at 3 divided by the vocabulary size.
Shuffle the American first-rank column
Keep every word, keep every number, keep every line of scoring code. Only permute which American frequency belongs to which word, inside each category. If shuffling the data does not hurt the prediction then the prediction was never using the data.
Both controls call the same judge(category, mode, word) the live round calls, with the same sealed guesses. Nothing is re-seeded, because the machine has no per-round memory to re-seed: its whole state is the norms, and that is the point of the second control.
A 2004 paper already told us which categories would travel.
The obvious dismissal is that 55% is a base rate dressed as mind reading. The interesting answer is that the base rate is not one number: it runs from 0.95 down to 0.00 across these 44 categories, and the variation is not noise. It was predicted in advance, in print, by a study that never saw this dataset.
Affinity scores suggest that 13 categories (12% of 105 categories) have roughly equivalent category responses across all four culture-by-age groups and are thus suitable for use in cross-cultural studies: time unit, internal organ, tree (part), season, color, face (part), farm animal, mathematical operation, metal, fruit, chemical element, human body (part), and wild animal. Yoon, Feinberg, Hu, Gutchess, Hedden, Chen, Jing, Cui & Park (2004), Psychology and Aging 19(3), 379-393, p. 390. Their comparison was American against Chinese, younger against older. It has nothing to do with Britain, or with 2018.
Six of those thirteen are among the 44 categories here: colour, fruit, metal, chemical element, unit of time and part of the body. If Yoon's list is measuring something real about which categories are culturally stable, those six should survive the Atlantic crossing better than the other 38. Press the button and find out; the split was fixed before the numbers were looked at, and it can lose.
The permutation test is the honest one here: the 861 responses are not 861 independent draws, because twenty of them share a category. So the test resamples whole categories, drawing six of the 44 at random and asking how often a random six beat the real six. That is a much weaker and much more truthful test than an interval computed over responses, and it is the number to read.
Every category, best to worst
The full table. Click a row to play that category. The from column is what the 2004 American norms nominate, translated into the British word where Banks and Connell judged them to be the same thing; a dash means the American word has no British counterpart in these norms at all, so it could never have scored.
| Category | Out-of-sample | American top three | British top three | Yoon |
|---|
Two categories score exactly zero. For family relationship the American norms nominate aunt, mom and cousin; the twenty British speakers said mother, brother and father. For weather the Americans nominate tornado, hurricane and rain; the British said rainy, sunny and cold, which is a grammatical difference as much as a meteorological one, and the matching rule is not clever enough to bridge it. That is a real failure of the instrument and it is left in the score.
Nobody has ever measured the thing this page is doing.
Every category production norm in existence is a timed fluency race. Banks and Connell gave their participants 60 seconds to say as many members as possible out loud. Van Overschelde and colleagues gave theirs 30 seconds to type. The quantity called "first-rank frequency" is the first item out of the gate in a burst, and a burst is shaped by retrieval dynamics that a single question does not have: people cluster, they warm up on an easy exemplar, they say the shortest word first because it is fastest to say.
Being asked, once, to think of a colour is not that task. It might well produce the same word most of the time. It might not. No dataset was found that measures the single-shot version, so this page cannot tell you the size of that gap, only that it exists and that it sits underneath every number here.
Who the twenty people were
Sixty native English speakers recruited from the Lancaster University student and staff population in 2017 and 2018, paid three pounds fifty, 46 of them female, mean age 21.72 years. The 117 categories were split into three lists of 39, so each category was seen by twenty people, not sixty. A first-rank percentage of 60% is twelve people out of twenty and has an exact 95% interval running from about 36% to 81%. Any single-category number on this page carries that width whether it is shown or not.
Who the Americans were
Van Overschelde, Rawson and Dunlosky collected at least 600 participants per category, mean 672, in a 2004 replication and expansion of Battig and Montague's 1969 norms. Their sample is far larger, and it is also American and more than a decade older than the British one it is being used to predict. Their published tables exclude any response produced by fewer than 5% of participants, which does not touch a top three but does mean their tail is not comparable to the British tail.
One more honest limit: the 2004 norms record first-rank frequency as a proportion, not a count, so the American side of every comparison is rounded to two decimals in the published table. Where two American words tie at the same rounded proportion the tie-break above does real work. That is measured in section three rather than assumed away.
Pick from the tail. It cannot follow you there.
The machine is spending one specific regularity: that a lot of people reach for the same few words. It has no other resource. Every category has a long tail of perfectly good members that no more than one or two people ever named first, and the moment you choose from there, the machine is finished. Here is the whole tail of the current category, with the exact number of people out of twenty who said each word first.
The way out, stated plainly
Feed it randomness and it loses. A coin, a dice roll, the fourth word rather than the first, a word from your own life rather than the category's centre of gravity: all of these beat it, and the second control above proves it by construction rather than by assertion. There is no adaptive component for you to fool; there is nothing here that learns you. It is a lookup table of what other people said, and the way to beat a lookup table is to not be in it.
The honest inverse
Being predictable here is not a flaw in you. It is the thing that makes a shared language possible. If the word you reach for when someone says "a colour" were unrelated to the word they reach for, the two of you could not use the word at all. The centre of gravity that this page is exploiting is the same centre of gravity that lets a stranger understand you on the first try. The machine is not reading your mind. It is reading the fact that you have one in common with several million other people.
What ran, what did not, and every free choice.
Predictions made before the data was opened, and how they did.
- P1 held. Predicted that the eight first-rank counts named in the build specification would reproduce exactly from the referential item table. All eight did: red 12, blue 5, apple 15, carrot 8, dog 11, hammer 15, monday 17, seconds 10, each out of twenty.
- P2 held, and it is the anchor. Predicted that all six of Banks and Connell's published rho values would reproduce from the comparison tables. All six did, to the two decimals they printed, including the sample sizes 1376, 595, 595, 201, 584, 267 and 54.
- P3 died, and it changed the headline. Predicted that the three hit rates handed down in the specification, 74.7% and 60.9% and 54.1%, would be three scorings of one instrument. They are not. Each was computed on a different footing: different word merging, different category set, different denominator. The fall from 75 to 54 was real but it was measuring three things at once. Recomputed on a single footing, the fall is 74.7 to 63.9 to 55.2, and the middle number moves by three points. The correction makes the page more interesting, not less: the cascade is now a clean measurement of leakage alone.
- P4 died. Predicted that the denominator would be 20 per category, giving 880 and 2340. It is not: 19 categories have fewer than twenty scoreable first responses, and the true totals are 861 and 2277, confirmed independently against the trial-level file. Using 880 would have invented 19 misses.
- P5 half died. Predicted from the specification that the per-category rate for insect was 0.15. It is 0.15 under an alphabetical tie-break and 0.35 under a tie-break by American production frequency, because bee and spider are tied at exactly 0.11 for third place. The number was not wrong, it was undetermined. That is now a measured sensitivity rather than a footnote.
- P6 held, unexpectedly well. Predicted that Yoon's 2004 whitelist would show some advantage on this completely different comparison. It shows a large one. Category-level significance is marginal with six categories, and the page prints the permutation p rather than a verdict.
- P7 not tested, deliberately. An Australian norming study from 2026 would have added a third scoreboard and a lovely result, opal and cricket and magpie appearing where the British and American norms have nothing. Its data sits behind a view-only token on a node whose licence field is empty. Redistributing it would have been a small theft, and citing a computation nobody can check would have been worse. It is dropped.
Primary sources, and what was checked against what.
Reproduce everything offline:
node research/the-word-you-were-going-to-say/build-norms.mjs
node research/the-word-you-were-going-to-say/verify-the-word-you-were-going-to-say.mjs