Some words only look like two words stuck together. CARPET is not a CAR plus a PET! Whether the two halves explain the word, or whether they just happen to spell it, is a question that a hand-made list answers badly, and on 20 October 2026 I filed an issue to replace the list with rules (#92). This post covers the rule engine that replaced it (#135), and the bug that turned up a few hours after it shipped (#161).
The hand list
Kompanion links two words that belong together, as in a phrase or a compound. Until 21 October it read a hand-made phrase list, and issue #92 says what was wrong with that: the list held chance splits, words that happen to break into two real words with no link in meaning. catKin works the other way. It takes a single word and breaks it into pieces, so what matters is that the pieces are readable, and the meaning does not have to follow.
| Mode | Rule | What a link looks like |
|---|---|---|
| Kompanion | Phrase | Two words that go together, such as HOT and DOG |
| catKin | Word | Pieces of one word, such as HAM and BURGER |
So a word like CARPET has to land in catKin, and the engine has to know why.
The evidence
The rules need something to read, so a prepare step (npm run joins:prepare) streams the English Wiktionary extract from kaikki.org and reads WordNet 3.1. It keeps only the evidence for candidate words and two-word terms, such as the word-formation templates and the two-word entries, and it drops every definition and all other text. The result is data/joins/evidence.json, which was 1,645,095 bytes, about 1.6 MB, when I measured it on 8 November, the same size as at #135.
Two small lists sit beside it. One holds curated two-word terms that the rules count even when WordNet does not list them, 3,184 of them at #135 and 5,907 on 8 November. A first pass picked those terms, and a person spot-checked a random sample, so not every term has had a review. The other holds 128 vouched pairs, which I chose by hand from the old list.
The data credits ship with the files. The Wiktionary content comes from its contributors under CC BY-SA 4.0, through the kaikki.org extract made with wiktextract, and the Wiktionary part of the evidence file stays under that licence. WordNet 3.1 comes from Princeton University under the WordNet licence.
The rules
The engine is a pure function. It takes a candidate split of a word W into A and B, plus the evidence, and it returns a verdict, the rule that decided it, and a short reason. The header of scripts/joins/rules.ts holds the rules, and this table is a plain version of them, with an example for each from a run on 8 November:
| Rule | Says | Example |
|---|---|---|
| R0 | I vouched for the pair, so it is Kompanion | SUPER + HERO |
| R1 | WordNet lists “A B” as a two-word entry, or the curated list holds it | HOT + DOG |
| R2 | Wiktionary marks W as a compound of A and B | BUTTER + CUP |
| R3 | A WordNet gloss of W, or a broader term, holds A or B | HONEY + BEE |
| R4 | W fails R1 to R3, and each side is a word, an affix or a chunk, so it is catKin | HAM + BURGER |
| R5 | A word that passes R0 to R3 on any cut never goes to catKin | SE + ASHORE |
| R6 | Rare words, unreadable sides and blocked words are rejected | HAMBURG + ER |
Some rules carry a gate, because a gloss can mislead. When Wiktionary traces a word to an older form that contradicts the cut, a single shared word in the gloss does not count. ISLAND looks like IS and LAND, and the linter, which is npm run joins -- WORD, prints the reason:
ISLAND
IS + LAND catKin R4 IS is a word and LAND is a word. Wiktionary traces ISLAND to an older form that does not start with IS, so one gloss word does not join it That is the first line of its output on 8 November. The point of the linter is that every verdict comes with a sentence, so a surprising join can be argued with.
Run over the whole word list on 8 November (npm run joins -- --report), the engine judged 95,136 candidates, which were 62,863 words and 32,273 two-word terms, and made these decisions:
| Rule | Splits decided |
|---|---|
| R0 | 203 |
| R1 | 33,550 |
| R2 | 1,514 |
| R3 | 168 |
| R4 | 22,060 |
| R5 | 437 |
| R6 | 31,705 |
R1 decides 33,550 of the 35,435 Kompanion splits. R3 decides the fewest, and the report lists its borderline cases, so that a person can check that the shared word is the meaning and not chance.
The old days
The rebuild was built to leave the past alone. #135 rebuilt all five modes for 706 dates, from 22 October 2026 to 26 September 2028, and I checked that no file for a date up to 21 October differed from main. All 1,343 tests passed at #135. Every daily up to 21 October stayed as players saw it, because a published day does not change.
That held for the boards, but one feature still read the word list, and it broke. That evening I filed issue #159. After a loss on the Rootlet of that day’s Kompanion daily, the board said that the answer was on its way, and then it said it could not recover the answer for this puzzle. The step behind “Show the answer” still read the current word list, and the new Kompanion list no longer held TOWN HALL, BOARD MEETING or BOARD MEMBER, which the older dailies had used. The live site showed it too. That was my bug.
A weekly sample of 2026 showed how wide it was. Before the fix, “Show the answer” failed for 77 of 106 Kompanion boards and 84 of 106 catKin boards, and every failure fell before 22 October. Klimber, sKion and Kalyx passed. The fix took the word lists out of that step, so that a later change to a list cannot break it again, and #161 merged at 20:46 Singapore time on 21 October, sixteen minutes after the issue. After the fix, a full sweep of every committed board passed in all five modes.
So the rule I took from it is that a published day must never depend on data that can change later.
Looking back
The rules replaced a hand-made list with an engine that gives each verdict a rule and a sentence. The first thing the new lists broke was the one place that still read them for old days, and the fix was to stop that place from reading any list.
The next post is about the text that you can share after a game, and how it shows a friend how you did.
Give it a go at kinopy.day and tell me which join surprised you!