Forgotten Languages: sources, word lists and methods

forgottenlanguages.org has published about 10,300 posts in some 150 invented "languages", many about UFOs, "Defense" programmes and contact. Most posts we could test are existing English texts, rendered word for word through a substitution table: folklore and myth from Project Gutenberg, the King James Bible, occult classics, two Dungeons & Dragons 3.5 supplements, arXiv physics papers, psychology journals, Wikipedia articles as they read on the day. The UFO layer sits mostly in the titles, in English passages FL writes itself, and in find-and-replace edits made to a source before encoding. Our conclusions.

What the numbers mean

A matched post has been lined up sentence by sentence with the English text it renders; the source is its translation. A matched post is often only partly matched: FL's own passages, quotes and reference lists are left out, so "sentences paired" matters more than the post count. About a third of the site is matched or readable; about a fifth of all words sit in sentences paired with their source. The rest is mostly a long tail of small languages (a handful of posts each) that open up once one of their posts has a source.

Every row can be checked: posts link to FL, and sources are cited as the original published work. If a match is wrong, please open an issue.

What Forgotten Languages is: our conclusions

After matching 3,329 posts to their sources and testing every method against controls, our answer (October 2026) is: Forgotten Languages is an elaborate body of writing built on top of existing texts, hidden behind codes that can be broken. Most of what the posts "say" is borrowed, and we can show from what. What is FL's own is the framing: the titles, the English passages it writes into the posts, the references, a recurring cast of characters and programmes, and, from about 2022, essays of its own. What that own material is — invention, allegory, something else — is not a question this method can answer, and we take no position on it.

The evidence in numbers

Where the matched posts come from, roughly: public-domain books from Project Gutenberg (folklore, myth, occult classics, literature; about 1,400 posts), two Dungeons & Dragons 3.5 supplements (about 750), Bibles (about 300), other books and web pages including Wikipedia (about 670), academic papers from journals and arXiv (about 110), theses, reports and government texts (a dozen), and translations made by the FL Discord group where no published original has been found (66). Matches run from FL's first year (2009) to its latest posts (2026).

How a Forgotten Languages post is made

1. An existing text is taken: a fairy tale, a chapter of a book, a paper, an encyclopedia article, a press release.
2. Sometimes its topic words are swapped before encoding ("terrorism" becomes "sightings", a particle collider becomes the "Queltron Collider"), so a real paper reads as UFO research. See Evidence.
3. It is written out word for word in one of FL's "languages". Most are fixed lookup tables (each English word has its own invented word); some respell English; a few render Spanish, Croatian or Serbian texts instead, and one writes English by sound in Chinese characters. The tables change over time (new tables in 2014, 2019 and 2020; from about 2024 several languages switch to a spelling with marked letters such as ż for "th").
4. FL's own material is added: English passages spliced into the text, invented references (including citations of FL's own posts and of works dated after the post), and a title that frames the borrowed text as UFO, "Defense" or contact material.
5. The same text is often published several times, in different languages or codes, under different titles.

Cassini Diskus, presented by FL as a "4-dimensional logical language", follows the same pattern: its typed symbol posts are Lord Dunsany's Fifty-One Tales (also identified independently by pilkingtonsbrain), a 2008 music-therapy thesis and a U.S. Air Force press release, each in a substitution code. See Cassini Diskus.

What is FL's own

The parts that cannot be found elsewhere are the titles, the English inserts, the references, and a body of essays from about 2022 on, many of which read like AI-assisted writing (a pattern pilkingtonsbrain first noted). Together they build a consistent and continuing setting: classified programmes and facilities ("DENIED"), named bodies such as SV17q and SV06n, the Denebians and Giselians, the recurring figure of Norea, probes, "Queltron" physics, and speculation about consciousness, artificial intelligence and the future of humanity. Our hand translations give a sample, and this material is where anyone who wants to go further should look.

Two checks we did run on it, for what they are worth: the one "event" title that matches real news (Holloman, 2014) carries the U.S. Air Force's own press release about a drone crash, encoded; and fourteen other event titles, checked against news for the same place and date, matched no reported incident. That is a small sample of titles, and it says nothing either way about the rest.

What we cannot say

Who writes Forgotten Languages, and why, is not something the texts reveal, and we do not speculate about it here. Nor can text-matching settle what FL's own material amounts to: our method can show that a passage comes from a 1915 book or a 2008 thesis, but it cannot tell you what the passages FL wrote itself are about, or whether anything behind them is real. Those questions are left open on purpose; this project is a set of sources and tools, not a verdict.

What remains open

• About 59% of posts with text are not matched yet, mostly in small languages with few or no pairs. Each new source opens up its language; see Help wanted.
• The Cassini Diskus picture posts (about 590): they use the symbols of one of FL's codes and were re-posted and extended under new titles, but whether they carry text at all is unresolved. See Cassini Diskus → Open question.
• Three typed symbol posts without a source (8027, 6898 "Dulce L7", 7187), and the 2026 "Beacon" pixel grids.
• FL's own essays are only partly translated; more translation would show more of what FL writes in its own voice.

Everything above can be checked: each matched post links to FL and to its cited source, and the methods and their controls are described under Method. If anything here is wrong, please open an issue.

Reskins: real papers with words swapped

Some technical "Defense" posts are real papers with their topic words replaced before encoding. Lined up with the original, the decoded post differs only in those words. Each example shows the original sentence and the FL version.

Cassini Diskus

"Cassini Diskus" is FL's name for posts written in small coloured geometric symbols: 711 posts from 2010 to 2026, many titled with map coordinates or "events" ("Phoenix Event", "Holloman AFB Downing Event"). FL has described it as "the 4-dimensional logical language that we call Cassini Diskus" and, in an email to a reader, as "our encoding of all the 100 most cited papers from CERN (2006-2012)".

An example

Each block is one symbol-word as FL wrote it (symbols and colours copied from the post); below it, the word Dunsany wrote at that place. A blank means the line-up skips that word.

How it was found

1. The symbol posts keep real spaces between "words", and every symbol carries its own colour.
2. A letter-by-letter codebreaker (the kind used on the Zodiac ciphers) broke test texts written in a similar cipher, but not Cassini Diskus: it is not a letter cipher.
3. Word repetition patterns: where the same symbol-word comes back, and how many words later, survives any word-for-word code. Compared with 9,500 English texts, eight posts matched the same book, Fifty-One Tales. The same posts reversed (the control) matched nothing consistently.
4. Lining the matched posts up word by word gave a dictionary of symbol-words. The dictionary placed the remaining posts in the book, which added more words, until no new posts were found.
5. Check: each post was decoded with a dictionary learned only from the other posts. 96% of its known symbol-words gave the word Dunsany wrote at that place.

How the code is built (October 2026)

• Two keys. Series A (April 2012 – July 2013) and series B (from 23 July 2013) use different symbol-words for the same English word: 93–97% of words in a series-A post use key 1, 82–92% in a series-B post use key 2.
• Punctuation is part of the word. An extra square at the end of a symbol-word marks the punctuation after it: dark-salmon ■ = full stop (92% of 883 cases), dark-slate-blue ■ = comma (93% of 1,025), dark-sea-green ■ = colon (96% of 96).
• It is built from spelling. Endings are regular (in key 1, -ed words end in the same two symbols 35 times out of 109; -ing, -ly and -s have their own endings), the first letter of a word predicts its first symbol, and consonants are kept more often than vowels. A model that learned letter-chunks from part of the dictionary read 39% (key 1) and 25% (key 2) of hidden words whose stem it had seen, and 10% of words with a new stem, against 0.1% for a shuffled control. So FL built the dictionary partly by rule and partly by hand.

What is not decoded, and why

All 711 posts FL labelled "Cassini Diskus" (2010–2026) were downloaded and sorted by content:

The decoded posts run through the book in order, twice (from April 2012 and from July 2013). Tales 42–50 ("Furrow-Maker" to "What We Have Come To") appear in no symbol post: FL seems to have stopped before the end of the book, about when the symbol posts turned into pictures (late 2013).

Why the 590 picture posts are not decoded.

• The symbols are tiny. Most are 6–7 pixels high, some 9–12. The pictures are about 435 pixels wide, and Blogger holds no larger originals (every size option was tried).
• Most are JPEG. Compression blurs a 6-pixel symbol enough that shapes such as ▣ ▦ ▨ ▩ can no longer be told apart. A symbol reader trained on the right font (Segoe UI Symbol, which Windows uses for these characters) recognised 6-pixel test symbols only 30% of the time (12-pixel: 56%).
• What the pictures are made of. The pictures are typed Unicode characters in Segoe UI Symbol (geometric shapes, plus 🍎 🍏 apples, ⧋ and ◫), in many colours and sizes, in lines of symbol groups: the same look as the decoded text posts, with a new set of symbols.
• The clean pictures do not behave like an encoded text. The 181 PNG pictures are lossless: every symbol is drawn pixel-exact and can be told apart, and their colours are exact HTML colour names, like the typed posts. The test: the decoded Dunsany posts were drawn as pictures in the same font and layout and run through the same reader. In those, 52% of multi-symbol words recur (2% when shuffled); in the real PNG pictures, 4% (0% shuffled). In the quality-100 JPEG pictures (2015–2018) the rendered Dunsany text still shows 23%, the real pictures 0%. A word-for-word code of a text repeats its common words constantly, so these pictures are either not a code of a text, or a code whose drawing FL varied every time. The low-quality JPEGs (2018–2023) cannot be tested: compression hides the pattern even in the rendered Dunsany control.
• Not used: AI upscaling. Upscalers (for example Upscayl / Real-ESRGAN) invent plausible detail; on 6-pixel symbols they can change one symbol into another, and differently each time, which would corrupt any count.

What would help. Higher-resolution copies of the pictures (screenshots, video frames, archives); any post or file that shows the same symbol text both typed and as a picture; or, as with Holloman, finding a picture post's source text from its title, date and other clues. The pictures, the symbol reader and the counts are on file and can be shared with anyone who wants to try.

The same picture text, posted again and extended (October 2026). Comparing the clean pictures symbol by symbol shows that FL re-posted some picture texts under new titles, usually weeks later and with more symbols added at the end, like a log or a text written in instalments. The re-posts are not copies of the same picture: the text was laid out again, with different line breaks and size. Found among the lossless PNG pictures (2014–2015):

Two more pairs share long stretches (5288/6190, 5296/6206), and 5957 also shares text with 5987. The same text under different "event" and coordinate titles means the titles do not describe the picture text. If the pictures encode a text, reading one member of such a family would read them all. No changing "fields" (such as a date) were found.

How the sharing works (corrected, October 2026). Lining pairs up run by run shows two different relationships. Usually the two pictures are one sequence lightly edited: the alignment offset stays within a few symbols from start to finish, with short insertions and deletions scattered through. In a few pictures (5744, 5957, and 7072) the repetition is internal: the picture carries the same block twice, which earlier looked like a later post adding material. Only three of 182 pictures do this. And the material is not confined to single posts: post 6206 (29 May 2014) shares four runs with post 5301 (26 July 2015) and two different runs with post 5296 (28 July 2015), while those two share nothing with each other — one picture's content distributed across two later ones more than a year afterwards. We found no case of blocks appearing in a genuinely different order. This line of questioning was prompted by xeniseni, who asked whether these are successive versions of one text or different arrangements of a shared set of elements.

A blur-tolerant comparison (symbols described by size, colour and rough outline, then lined up allowing mismatches, as DNA sequences are compared) extends this to the JPEG pictures. Tested on the known PNG pairs re-saved as JPEG, it finds them easily at quality 100, weakly at quality 75 and not at all at quality 60, so low-quality pictures can hide more families. Result: 16 families covering 41 posts. The largest is one text across nine posts from May 2014 to August 2015 (6225, 6206, 6203 "Dillon Event", 6190, 5784, 5301, 5296 "Paintsville, KY", 5289, 5288). Four families are JPEG pictures from 2016, 2017, 2018 and 2020 (4572/4554, 4078/4076, 3586/3584, 2605/2604), so the re-posting continued after the PNG years.

Open question: are the picture posts text?

This is not settled, and we would welcome other eyes on it. What the tests show (October 2026, each with a control):

• They use a known code's symbols. The 2013–2015 pictures are drawn in the palette and with the commonest symbols of the last typed symbol posts, 6898 "(Dulce L7)" and 7187, which use one symbol per English word (◥ pink is the commonest symbol in both, as "the" is the commonest English word). 90–97% of the picture symbols are in that code's colours.
• Counted one symbol at a time, they look like words. Commonest symbol 8.3% (English "the" 7.0%); variety of symbols 0.46 per symbol (English words 0.46); symbols recurring within 63 places 42% (English 39%).
• Their order does not look like English word order in any simple reading direction. A word-substitution solver that reads the common words of a coded English control text gains much more from real word order than from shuffled order (+0.64); on the pictures the gain is small (+0.11), and reading right to left, bottom up, alternately or by columns does not change that. Repeated symbols also sit closer together than repeated words do in English (20 places apart against 33).
• No source has been found. A key-free search of every PNG picture against about 11,000 texts and every FL text post found nothing above the reversed-order control (the same search finds a coded Dunsany test passage easily).
• Two other readings have now been tested, and neither fits. FL has one language that writes English by sound in Chinese characters, so the pictures might spell sounds rather than words; and FL renders Spanish texts in several languages, so the text might not be English at all. A sound system draws on a small closed set of units and repeats it hard: that Chinese-character language uses 509 distinct characters across 94,000 units, its commonest covering 4.7% and its top twenty 40% of the whole text. Within a single picture the symbols are far more varied than that — roughly 200 distinct shapes across 400 positions — which is the profile of words, not syllables. For Spanish, the word-order test above was repeated with a Spanish model built from three million words of Spanish pages: coded Spanish gains +0.60 from real order, so the test works, but the pictures gain +0.10, the same as under English. Both results depend on our cutting the pictures into symbols correctly, which is the weakest joint in all of this.

Possible explanations, none confirmed: the pictures are generated from the code's vocabulary rather than encoded from a text; they encode a text in an order we have not found (a layout rule, a scrambling key, a different reading path); they encode a language we have not tried (English and Spanish have been tested, and a sound-based reading looks unlikely); or they encode something that is not a text. Typed posts 6898 and 7187 themselves do behave like English text, so finding their source would give the code's dictionary and could settle the question.

Open problems – help wanted

Four typed symbol posts are not Dunsany; one is now decoded (8773). The other source texts would give new dictionaries; if you have a lead or a candidate text, please open an issue.

The titles

Many Cassini Diskus titles name an "event" or give map coordinates. One is real news: "Holloman AFB Downing Event" (8 February 2014) is a different symbol code, and its text is the U.S. Air Force press release on the unmanned QF-4 target drone that crashed near Holloman the day before (matched clause by clause; the release is public domain). Fourteen other event titles were checked against news from the same place and date (Abu Dhabi, Incirlik, Dry Lake Valley, Tuapse, Loughton, Ottumwa, Antalya, Stamford and others): no real incident. Some reuse well-known UFO stories (Phoenix Lights 1997, Aztec 1948, the Philadelphia Experiment, Pine Gap). The coordinates are real places; a few are star positions: Tau Ceti, and Deneb Algedi (δ Capricorni), the "Deneb-Algedian" probes FL wrote about in an email.

Neither FL description fits what the decoded posts contain: there are no CERN papers and no logical formulae, only Dunsany's fables. This is the same pattern as the rest of the site: an existing text, re-encoded, under FL's own titles.

Thanks to pilkingtonsbrain, who independently tied the same symbol posts to Dunsany's Fifty-One Tales; to a 2024 Reddit thread on decoding the Cassini Diskus, which noted that the symbols group into repeating "words"; and to the collectors of the old FL Group forum archive. Data: cassini_posts.csv (every symbol post and its tale), cassini_codebook.csv (the dictionary). Source text: Fifty-One Tales at Project Gutenberg.

All symbol posts

Hand translations

English versions of FL posts, made from our word-by-word decoding. Most are FL's own writing from 2023–26. Posts are listed from most to least readable. Click a post to open its translation.

For an automatic word-by-word decoding of every FL post, see pilkingtonsbrain's Forgotten Languages, Decoded. Two of the posts here (Signs of otherness, The Great Cooling) were translated earlier in the FL Discord group by Purplexed1.

Also as a file: hand_translations.md

Posts

Matched posts, posts whose source is known but not yet in hand, and well-decoded posts still without a source. Search by title, post number, language or source.

Test = an independent check of each match: confirmed = the post's sentence lengths track the source (no dictionary used) and its decoded words land in the matched source sentences far more often than in other texts; likely = one of the two (usually the post is too short for the length test). On deliberately wrong pairings no match was confirmed. Covers = how much of the post the source accounts for: many posts mix a borrowed text with FL's own writing, so a confirmed source can still cover only part of a post.

Identified sources

original work identified = the citation of the original work was checked against a catalogue, Crossref, arXiv or the title page (this is about the citation only; how well each post matches is shown on the Posts tab); text from a copy = original known, text came from a mirror; web page only = original not confirmed; Discord group translation = the group's own English of a post.

Languages

"Readable" = median share of words we can decode in the language's unmatched posts. Notes give aliases and later stages (a language that switched to a new word list at a date).

Word lookup

Paste FL text (a word or whole sentences) and pick its language. Each word is looked up in that language's list; [?word] = not in the list. Readings with a slash were both seen. This is a plain word-for-word lookup: no guessing is added.

Help wanted

The quickest progress comes from finding source texts. A PDF, page link or library copy of any of these can be matched and tested, and teaches its language new words. Four Cassini Diskus symbol posts with their own codes are listed separately under Cassini Diskus → Open problems (one decoded, a Dulce text wanted).

Sources known or suspected, text not yet in hand

Decoded well, source unknown

Search a distinctive phrase from the decoded opening. [?word] = unknown, [word?] = doubtful.

Using this with an AI assistant

Credits