Ladder project · 13 November 2026
Nine-Letter Word Ladders
In 2000 Rex Gooch set out what an ideal word ladder between nine-letter words would be, and showed how nearly one could be built from the big dictionaries. We ran his rules over modern word lists, his own dictionary files and a 2.6-million-string corpus. Familiar words alone give nothing. With a small set of judged words we get four ideal ladders and three near-ideal ladders between heterograms that are not in his table. No nine-letter ladder of listed words changes its letters in order. At eight letters one does, on two obsolete words; at nine it takes three coinages and four old spellings, and a wider search over 35 million made words, compounds and corpus strings finds nothing better.
Gooch's rules
A word ladder changes one letter at a time, every step a word. In "Near-Ideal Nine-Letter Word Ladders" (Word Ways 33, 2000, pp. 209–211) Gooch listed six properties a ladder between N-letter words might have:
- it takes the minimum number of steps, N;
- the last word differs from the first in every position;
- the first and last words are heterograms: no letter repeated;
- the letters are changed in sequence, first to last or last to first;
- every word is recognisable;
- the first and last words are related, in meaning or as wordplay (reversals, say).
He called a ladder with properties 1 and 2 an ideal ladder. His second table held ladders between heterograms with one letter left unchanged: eight positions differ, in at most nine steps. We call those near, and near* when eight steps suffice. He found no nine-letter ladder with property 4, and his ladders lean on obsolete forms from the Oxford English Dictionary.
The search
The program takes every word as a start and searches outwards, one letter at a time, to nine steps. For each pair of words it records the shortest ladder, how many shortest ladders there are, and the one whose least familiar word is most familiar. It flags heterogram ends, ladders that change their letters in order, and reversals. A full pass over a few hundred thousand nine-letter words takes seconds.
What matters is the word list, so the lists were added in tiers, most trusted first:
- Trusted words: the two more familiar grades of a graded word list, plus the tournament and reference lists (ENABLE, Webster's Second, the Scrabble lists): 94,379 nine-letter words.
- Gooch's own dictionary file, which records the dictionaries each word is in: 51,784 nine-letter words in two or more of his dictionaries, 166,356 in exactly one, 45,176 that are his own inflected forms.
- A corpus: 2,593,711 distinct nine-letter strings from a word-frequency database, split into those in Wiktionary (78,521), those frequent in Google Books, and the rest.
Words from the looser tiers were not taken on trust. Each word a ladder depended on was looked up in the local dictionaries, then judged twice, by a language model and by a second reviewer, with a one-line definition as the test; the stricter verdict stood. Three words on which the judges disagreed were allowed by a human ruling.
What each word list gives
| Nine-letter words used | Ideal | Near* | Near |
|---|---|---|---|
| Trusted words only | 0 | 0 | 1 |
| + words found in the dictionaries or judged in the first two passes (20 words) | 3 | 0 | 1 |
| + judged Wiktionary words (14 words) | 4 | 0 | 2 |
| + the three words allowed by ruling | 7 | 4 | 10 |
| Unjudged tiers, for comparison | |||
| Trusted + Gooch's file, two or more dictionaries | 3 | 0 | 1 |
| … + words in one dictionary | 100 | 19 | 51 |
| … + his own inflected forms | 184 | 43 | 93 |
| Trusted + all of Wiktionary | 17 | 5 | 11 |
The one ladder the trusted words give is already in Gooch's table: UNDERLOFT to UPLOCKING. The three ideal ladders of the second row are his too: RIDGLINGS to SANDPAPER, and IMPRESSOR to UNTAINTED or UNSAINTED. The search for missing rungs landed on the same obsolete words he used, such as IMPRENTED.
The new ladders
Words in this colour were judged marginal: real, but rare, obsolete or a plural seldom seen. Words in bold were allowed by ruling after the judges disagreed. The rest are in the trusted lists.
Ideal: nine steps, every letter changed
- PEDALINGS
- PEDDLINGS
- PADDLINGS
- SADDLINGS
- SANDLINGS
- SANDLINES
- SANDPINES
- SANDPIPES
- SANDPIPER
- SANDPAPER
- BROODINGS
- BLOODINGS
- BLONDINGS
- BLONDINES
- BLONDINED
- BLONDENED
- BLINDENED
- BLINDERED
- BLINKERED
- CLINKERED
- BROOMINGS
- BLOOMINGS
- BLOODINGS
- BLONDINGS
- BLONDINES
- BLONDINED
- BLONDENED
- BLINDENED
- BLINDERED
- FLINDERED
- BLINDERED
- BLINDENED
- BLONDENED
- BLONDINED
- BLONDINES
- BLONDINGS
- BLOODINGS
- BROODINGS
- BROOMINGS
- GROOMINGS
Near*: heterogram ends, eight letters changed in eight steps
- CHORDLIKE
- CHORELIKE
- SHORELIKE
- SHORELINE
- SHORELING
- SHOVELING
- SHAVELING
- SLAVELING
- SLAVERING
- CHAVELING
- SHAVELING
- SHOVELING
- SHORELING
- SHORELINE
- SHORELIKE
- SPORELIKE
- SPORTLIKE
- SPORTBIKE
- BORINATES
- BORONATES
- CORONATES
- CORONITES
- CORONISES
- COLONISES
- COLORISES
- CALORISES
- CALORISED
The fourth near* ladder, CLAVERING to SHORTLIFE through SHORELING, is Gooch's. Nine more pairs of heterograms are joined in nine steps, starting from CHAPELING, CHORDLIKE, CLAVERING, PEDALINGS and BORINATES; one is PEDALINGS to SANDVIPER, which follows the first ladder above to its last rung.
An ideal ladder between heterograms, from the OED
None of Gooch's ladders has properties 1, 2 and 3 together. His own dictionary file holds one that does, if words found in a single dictionary are allowed:
- CLAVERING
- SLAVERING
- SLAVELING
- SHAVELING
- SHOVELING
- SHORELING
- SHORTLING
- SHORTLIND
- SHORTLAND
- SHORTLAID
We put its three unjudged rungs to the judges. SHORTLING passed as marginal: an old spelling of shorling, the skin of a shorn sheep, which Gooch himself names as a possible rung. SHORTLAID passed as marginal: short-laid, said of rope laid with a short, hard twist. SHORTLIND failed with both at first: it is a hyphenated string in an OED quotation, not a headword, and neither judge could say what it meant. Gooch's notes on his file settled it. He took 124,968 hyphenated words from the OED's quotations, dropping hyphens and apostrophes, and the file has WELLLIND beside WELLLINED, and SHORTLIVD and ILLLOOKD: well-lin'd, short-liv'd, ill-look'd. So SHORTLIND is short-lin'd, an old spelling of short-lined. We infer that from its neighbours and have not seen the quotation. On that evidence both judges passed it as marginal, the same standing as IMPRENTED in Gooch's own ladder. The pair has no other ladder of nine steps, so this one stands or falls with SHORTLIND, and as judged it stands: an ideal ladder between heterograms, on one word allowed by ruling and three marginal ones.
The words they rest on
Every new ladder hangs on a few uncommon words. These are the ones that matter, with the verdicts as given.
| Word | Verdict | What it is |
|---|---|---|
| PEDALINGS | marginal | plural of the gerund pedaling; rare |
| SADDLINGS | marginal | plural of saddling; also in Gooch's own ladder |
| BLONDENED | marginal | blonden, to make or become blond; rare |
| BLINDERED | marginal | wearing blinders; the first judge called it archaic, the second current American |
| CHORELIKE, SHORELIKE, SPORTLIKE | marginal | regular -like formations, in Wiktionary; rare |
| BORONATES, CORONITES | accept | chemistry and petrology |
| SHORELING | allowed | OED variant of shorling, a shorn sheep or its skin; the first judge rejected it |
| BLINDENED | allowed | blinden, an archaic verb, to make blind; the first judge rejected it |
| BORINATES | allowed | borinate, an ester of a borinic acid; the first judge did not know the word |
Seven other disputed words stayed out: BROOKINGS, SPLITTIES, SPLITTEES, SENDLINGS, BLICKERED, CHLORINEY and UNDERRANG. The first judge also passed a good many machine-made forms in the earlier rounds (FOILERIES, TONSILEST), which is why two judges were used and the stricter one believed.
What the corpus could not give
Strings that are merely frequent in scanned books are no use as rungs. Adding the 71,843 most frequent ones produced 404 ideal ladders, among them the first that change their letters in order. Every one we looked at was a ten-letter word losing each letter in turn:
- APARTMENT
- DPARTMENT
- DEARTMENT
- DEPRTMENT
- DEPATMENT
- DEPARMENT
- DEPARTENT
- DEPARTMNT
- DEPARTMET
- DEPARTMEN
A filter now drops any string that is a trusted word with a letter missing, doubled, put in front, or two neighbours swapped. It removed about 5,200 of those strings and every false sequential ladder. What the frequent strings then add is Latin (IMPRESSUS, IMPRESSIS). The next 219,776 strings still give thousands of ladders, but the ones sampled are misprints (OPERSTLNG) and surnames (MUSKOWITZ), and that tier was not judged.
Shorter words, for scale
Nine letters is where the game becomes hard. With the same trusted words:
| Letters | Words | Ideal | Near* | Near | Notes |
|---|---|---|---|---|---|
| 9 | 94,379 | 0 | 0 | 1 | |
| 8 | 81,996 | 595 | 168 | 700 | 17 ideal between heterograms; none in sequence until two OED words are added (below) |
| 7 | 62,112 | 131,873 | 45,193 | 184,511 | 330 ideal in sequence, none of them between heterograms |
At seven letters property 4 is within reach: CLASSIC-CLASSIS-CLASSES-CLASHES-CLACHES-CLOCHES-COOCHES-MOOCHES changes its letters in order from last to first. At eight, COUGHERS-ROUGHERS-ROUGHENS-ROUGHINS-ROUGHING-ROUCHING-ROACHING-REACHING-REACTING is ideal between heterograms.
Letters changed in order
Property 4 needs its own search. A ladder that changes its letters in order is fixed by its two ends: rung k is the first k letters of the last word followed by the rest of the first. So there is nothing to choose; either all the in-between strings are words or they are not. We walked every trusted word forward one position at a time, letting a rung be a trusted word, a word from the looser tiers, or, once or twice per ladder, a string in no list at all, to be put to the judges.
Nine letters: none among listed words
With every local list admitted there are sequential ladders, but each one passes through misprints from the corpus. With real words and up to two invented rungs, only two ladders come within one rung, and the rung is impossible:
- UNPAINTED
- INPAINTED
- IMPAINTED
- IMPRINTED
- IMPRENTED
- IMPRESTED
- IMPRESSED
- IMPRESSOD ✗
- IMPRESSOR
- TWILLINGS
- SWILLINGS
- SHILLINGS
- SHELLINGS
- SHEALINGS
- SHEARINGS
- SHEARLNGS ✗
- SHEARLEGS
- SHEARLESS
The ladders one invented rung short need 45 different strings, such as UNDERSOIG between UNDERSONG and UNDERSOIL. All were rejected. The wall here is not vocabulary: the missing rung is one exact string, and it is rarely shaped like a word. A wider pass, with end words from any list and 382,774 words in all, found no ideal sequential ladder either.
Nine letters: one ladder, mostly made
The last resort of the method is to propose plausible unlisted words. We made them by rule from trusted stems: any noun with -like or -less, any verb with the old -est and -eth endings that Gooch himself used, and so on, 111,629 nine-letter forms. With those as candidates there are 22 ideal sequential ladders, 18 of them one family. The best of the family:
- JUDGELIKE
- BUDGELIKE
- BADGELIKE
- BARGELIKE
- BARBELIKE
- BARBALIKE
- BARBARIKE
- BARBARYKE
- BARBARYNE
- BARBARYNS
It should be read for what it is. JUDGELIKE and BARGELIKE are dictionary words and BADGELIKE is in Wiktionary. Three rungs are coinages in -like: on budge (lambskin fur), barbe (a Barbary horse, or part of a headdress) and barba (the beard, in anatomy). The last four are obsolete spellings recorded in the OED, two of barbaric and two of a noun barbaryn whose meaning we have not checked; the first judge rejected all four until shown the entries in Gooch's file. BARBALIKE was rejected by the first judge twice, as a formation nobody would use, and stands only by ruling. Every ladder of the family passes through it. So property 4 can be met at nine letters, but only by giving up most of property 5.
Nine letters: more rules and compounds, nothing better
Could a wider net do better? We tried three more sources of candidate words:
- More rules. Further suffixes and prefixes on trusted stems (-ment, -age, -ism, -ist, -ful, -ship, -wise, -ise, de-, dis-, fore-, semi- and others): 734,088 nine-letter forms.
- Compounds found in the corpus. 19,032 written there with a space or hyphen, every part a trusted word, and 339,206 written solid as two familiar words. The second set is noisy and lets misprints in.
- Every join of two familiar words of three or more letters, attested or not: 30,973,814 nine-letter strings. This was Eric's suggestion, and it is feasible only because the constraints are so tight.
Listing ladders one by one does not scale to 32 million words, so the search was rewritten. What can follow a rung depends only on the rung and the position reached, so the cheapest ladder from each word can be computed position by position from the last back to the first. That takes about a minute for all 32,319,089 words, and it is exact: it finds the best ladder, not merely a good one.
With everything admitted, 4,766,032 words start an ideal sequential ladder, and the cheapest need only four rungs from outside the trusted lists. There are 93 of those. One of them:
- WHALELIKE
- SHALELIKE
- STALELIKE
- STOLELIKE
- STORELIKE
- STORYLIKE
- STORYPIKE ✗
- STORYPOKE ✗
- STORYPOLE
- STORYPOLL ✗
A cheap ladder is worth something only if its rungs are words, so the search and the judges were put in a loop: find the cheapest ladders, send their unjudged rungs to the first judge, strike out what it rejects, and search again. The loop ran 30 rounds and judged 21,805 strings. It rejected 20,165 of them, and with those gone no ladder is left with fewer than eight rungs from outside the trusted lists. Every ladder with four, five, six or seven such rungs lost at least one of them: STORYPIKE, SNIFFLINT, BARRELAGS, BALLATONE, PRICELINT and the like among the glued pairs, and misprints such as SHORTHOND, COLLECION and DIRECIONS among the solid corpus compounds.
The first judge needed a check of its own. For a glued pair it will often invent a meaning: it passed SPICKDUNE as "a dune characterized by its neat appearance" and PINKSLEDS as sleds that are pink. Twice the loop stopped on a ladder it had passed, such as one ending STOOLBIKE-STOOLBAKE-STOOLBALE-STOOLBALL. So every string it passed went to a second, stricter question (have you actually met this exact string as a word?), which threw out about four in five of them, and the second judge reviewed the first few hundred by hand.
The judges also erred the other way, rejecting real words they did not know, such as TRACKLETS, a term from particle physics. So every string was also looked up in the word database, and one found in a dictionary (the OED, Webster's Second, Chambers and others, through Gooch's own source codes) or in Wiktionary passed on that evidence whatever the judges had said. That reinstated 776 strings. The loop was run again with them, and the result did not change.
A last, deeper pass widened the pool to 35 million strings. It added 1.7 million forms made by the same rules from the dictionary words in Gooch's file (not just the familiar ones), and 1.8 million strings from the corpus that had no dictionary behind them, after dropping broken forms of longer words. The wider pool does hold cheaper ladders. One family runs through seven listed words in a row, UPLOCKING-UNLOCKING-UNDOCKING-UNDECKING-UNDERKING-UNDERSING-UNDERSONG, and needs only three more rungs, but those are strings like VPLOCKING and UNDERSORG. This time the loop ran until the cheapest ladder left in the pool had every rung passed. After 19 more rounds, with 56,047 strings judged in all and 52,553 rejected, that ladder was JUDGELIKE to BARBARYNS again.
The newest source is a count of words in ordinary web prose, from 100,010 files of the Common Crawl: every nine-letter string used, mostly in lower case, on at least three different sites. It added 243,000 strings the pool did not have. What they brought to the cheap ladders was misspellings (SCIENTSTS, REPREATED, REPLICTED), and after seven more rounds, 68,029 strings judged and 63,907 rejected, the answer was the same.
We also dropped the two noisy sets (the glued joins and the solid corpus compounds) and searched the 1,052,833 words that remain, all of them formations a judge can at least define. Only 153 words start an ideal sequential ladder there, and the cheapest is the one already shown, JUDGELIKE to BARBARYNS. The runners-up, such as a family through STEEPWISE-STOEPWISE-STOMPWISE, are worse. Either way the answer is the same: fewer made-up rungs can be had only with compounds that mean nothing.
Eight letters: an ideal sequential ladder
So we moved down a letter, where the trusted words also give no ideal sequential ladder, and added Gooch's dictionary file. One ladder needs only two words from it:
- WHITTLES
- SHITTLES
- SPITTLES
- SPATTLES
- SPARTLES
- SPARKLES
- SPARKIES
- SPARKINS
- SPARKING
- OUTLINES
- CUTLINES
- CATLINES
- CARLINES
- CARDINES
- CARDITES
- CARDITIS
- CARDITIC
The two marginal words are both obsolete and both in the OED. SHITTLES is the plural of shittle, an old form of shuttle. SPARKINS is the plural of sparkin, a small spark or particle of fire, which the OED records only in Middle English (1150–1500). The first judge rejected SPARKINS outright, and a web search says flatly that it is not a valid plural. It was passed as marginal only after the evidence was put: the OED headword, and the fact that Gooch's own file lists SPARKINS under the same code as IMPRENTED, a rung of one of his printed ladders. That is a verdict reached by argument, and the reader may weigh it accordingly.
Gooch's "Sequential Snakes and Ladders" (Word Ways 33, May 2000) counted ladders that go there and back in sequence and found none at seven letters; his eight-letter article calls sequential replacement "perhaps the only element lacking". Neither prints a one-way sequential ladder of eight letters. The other properties are still missing here: neither end is a heterogram, and whittling has little to do with sparking.
Checked against Gooch's printed tables
Gooch prints ten ideal ladders and eight ladders between heterograms, with notes on variants, and says there are about fifty ideal ladders in all. We compared ours with his, ladder by ladder.
His ladders, in our search
Once every tier of his dictionary file is admitted, the search finds the end words of 17 of his 18 printed ladders. The one it misses is BUTT-SLING to CANCERATE, which needs CANCERINE, a word he took from the 1932 Webster and which is not in the tiers we used. With only the judged words, four of his ladders appear:
- RIDGLINGS to SANDPAPER, rung for rung as he prints it.
- UNDERLOFT to UPLOCKING, rung for rung.
- CLAVERING to SHORT-LIFE, in the eight-step form he marks with an asterisk (without GLAVERING).
- IMPRESSOR to UNTAINTED and to UNSAINTED, both among the variants he lists, but by a different route: ours goes IMPRINTED-IMPAINTED-INPAINTED-UNPAINTED, his IMPRINTED-INPRINTED-UNPRINTED.
Our ladders, in his tables
None of the ladders shown above as new is in his nine-letter tables: not PEDALINGS to SANDPAPER, not the three BLONDENED ladders, not CHORDLIKE to SLAVERING, CHAVELING to SPORTBIKE or BORINATES to CALORISED, and not CLAVERING to SHORTLAID. Three of them owe him something all the same:
- The PEDALINGS ladder is his eight-letter ladder in the plural. Three months later, in "Near-Ideal Eight-Letter Word Ladders" (Word Ways 33, pp. 284–286), he printed PEDALING-PEDDLING-PADDLING-SADDLING-SANDLING-SAND-LINE-SANDLIKE-SAND-PIKE-SAND-POKE. Ours follows it for six rungs and then joins his SANDPAPER ladder.
- The SHORELING ladders are built on his spine. SLAVERING-SLAVELING-SHAVELING-SHOVELING-SHORELING-SHORELINE is the middle of his CLAVERING ladder; CHORDLIKE and SPORTBIKE are new ends for it.
- CLAVERING to SHORTLAID uses SHORTLING, which he names as a variant rung. In the eight-letter article he says ideal ladders between heterograms appear "for the first time" at eight letters, so he had none at nine.
In a later note (Word Ways 34, August 2001, p. 189) he added a seventh condition: every word in an ideal ladder should be a heterogram. None of ours meets it; PEDDLINGS, CHORELIKE and BORONATES each repeat a letter.
Limits
- No sequential and no reversal ladder of nine letters turned up in any tier that holds listed words. The one sequential ladder shown needs three coined words, one of them allowed against a judge.
- The coined words came from rules, two-part compounds and the corpus. Within those the JUDGELIKE ladder is the best there is, since that search was exact. Three-part compounds, phrases and formations outside the rules were not tried. The 63,907 rejections that cleared the cheaper ladders were made by a language model and were not reviewed one by one. Each was checked against the dictionaries and Wiktionary in the word database, but a real word found in none of those and wrongly rejected could still hide a better ladder.
- The eight-letter search was narrower than the nine-letter one: trusted words plus Gooch's file, without Wiktionary or the corpus. Four more ideal sequential ladders there need three to six one-dictionary words each and were not judged.
- Property 6 was not judged. Nothing here checks whether the end words are related.
- The comparison with his tables was made from a scanned text of the article, not the printed page, and he prints about a fifth of his "about fifty" ideal ladders in full; the rest are notes such as "7 more ladders by using …".
- The verdicts are judgments. A stricter editor would drop the marginal plurals, and with them every new ladder; a looser one would take the seven words left out and gain a few more.
- The rest of the corpus was not run: 2.2 million rare strings, where a real word would be hard to tell from noise.
Sources and method
Rex Gooch, "Near-Ideal Nine-Letter Word Ladders", Word Ways 33 (2000), pp. 209–211. Word lists: a graded word list (grades 0 and 1), ENABLE, Webster's Second, the OSPD/OSW, NWL2023 and CSW24 lists; Gooch's dictionary file with its per-dictionary codes; a word-frequency database with Wiktionary, Google Books and Usenet evidence. Programs, all C++: a breadth-first ladder search, a missing-rung search that allows one two-letter gap, tools to split the dictionary file and the corpus into tiers, the misprint filter, generators for rule-made words and compounds, and a dynamic-programming search for the cheapest sequential ladder. Judging: a language model as first judge and a second reviewer, the stricter verdict final, every verdict cached with its definition.