# Briefing for AI assistants: reading Forgotten Languages posts You are helping someone read posts from forgottenlanguages.org ("FL"). This file sums up what one long-running project learned, so you can use its data instead of starting from scratch. All data files are linked at the end; fetch the ones you need. Everything here is checkable: posts link to FL, sources are cited as published works. ## 1. What FL posts are - About 10,300 posts (2008-2026) in some 150 invented "languages" (FL calls them cryptolects): Weddag, Niriden, Yid, Aylid, Nymma, Ladd, Dirya, Hawa, Laari, Dediaalif and a long tail of small ones. - **Most posts that could be tested are existing English texts rendered word for word** through a per-language word list: Gutenberg folklore and myth, the King James Bible, Theosophy and occult books, two Dungeons & Dragons 3.5 supplements (*Book of Vile Darkness*, *Book of Exalted Deeds*), arXiv and journal papers, Wikipedia articles as they read on the post's date, government reports. - 3,154 posts (39%) are matched to their source. For those, the **source text is the translation**: read the source, not a decoding. - The UFO / "Defense" layer sits mostly in (a) the post titles, (b) English passages FL writes itself, and (c) **reskins**: a real paper with its topic words swapped before encoding (terrorism -> sightings, ants -> Sol-3 humans, CLIC -> Queltron Collider, lunar -> exoplanet, 2012 -> 2030). The swaps are the FL-specific content. - Many posts are **collages**: a stretch of one source, a stretch of another, FL's own English paragraphs, quotes and a reference list copied from the field. - One text is often published in several languages ("twins"). Translate or source one and the others open up. ## 2. How to translate a post (do these in order) 1. **Identify the language.** The post's label usually names it; otherwise compare its common words with the word lists. `4_languages.csv` lists aliases (Dialect B = Weddag, Dialect C = Niriden, Nhue = Ladd cryptolect, Niriden2 = Niriden) and later stages. 2. **Check whether it is already matched.** Look the post up in `1_post_sources.csv` (by post_id or URL). If it is there, the cited source is what the post says. Note the `sentences_paired` count: a post can be only partly matched (collage), so the rest may need decoding. 3. **Check twins.** `5_twins.csv`: if another post in its group is matched, the same source applies. 4. **Check the open lists.** `2_still_to_find.csv` (source known or suspected, text missing) and `6_hunting_list.csv` (already decoded, source unknown, with the decoded opening). 5. **Decode word for word** with the language's word list (`word_lists/.csv`: fl_word, english, how_learned). Match whole words, lower-case; try with and without punctuation and accents. - Keep the FL word order: it follows the English source's order. - Mark every unknown word as `[?flword]`. **Never fill an unknown from imagination**; if you guess from context, mark it `[guess?]` and say so. - `a/b` in the english column = two readings seen; choose by context. - Add only small function words in brackets if needed for readability. 6. **Recognise the source.** A decoded opening with 60%+ known words usually reads as recognisable prose. Search a distinctive 6-10 word phrase from it (exact-phrase search), or by the topic plus the post's English quotes. Papers: arXiv, Semantic Scholar, PubMed Central. Books: Project Gutenberg, Internet Archive. Wikipedia: the article's revision as of the post date. If found, compare sentence by sentence and list any swapped words: those are the reskin. 7. **Report honestly**: how much was decoded, what is the source's content, what is FL's addition. ## 3. Things that will trip you up - **English inside posts.** FL leaves English quotes, titles, epigraphs and reference lists in its posts. Those are not FL words. FL also writes whole English "interview" or "Defense report" paragraphs; those are FL's own. - **Word-list errors.** Lists are learned automatically; roughly 1 in 10 words has two readings, some of them alignment slips. Readings learned from one post can be wrong. "how_learned" = "statistical model" or "twin transfer" are slightly less reliable than "mined from pairs". - **Very common words** (the, of, and) often have several FL forms in one language; that is normal. - **Stages.** Some languages changed their word list at a date. Niriden switched in December 2014: later posts ("Late Niriden") use a new list (ny = the, dy = of, le = to, ed = is, ad = they, pock = he, deayrt = said, rie = I/in). Dediaalif changed around 2020 ("Late Dediaalif"); Ladd 2019 posts use accented forms. Use the stage's own list when it exists. - **Respelling languages vs arbitrary lists.** Ladd, Niriden and Weddag words loosely resemble the English spelling ("I Puk Aff I Sysemys Aff Enok" = "The Book of the Secrets of Enoch"), but too loosely to decode unseen words by rule. Yid, Aylid, Hawa and Dirya words are arbitrary: only the list works. - **Grammar pieces.** Dirya attaches aţ- "of" (aţfw "of the"); Nymma and Dirya have suffixes. Split only when the remainder is a listed word. - **Other alphabets.** Some languages are partly written in Arabic (incl. Kurdish/Uyghur-style letters), Hebrew, Cyrillic, Greek or Devanagari letters: transliterate letter by letter to Latin, then look the word up (try vowel variants). One language is written in Chinese characters: it spells English BY SOUND, 1-4 space-separated character groups per English word (worked out by pilkingtonsbrain, "Forgotten Languages, Decoded"). His sound table is in the word list for that language (how_learned = "pilkingtonsbrain's lists"). - **Spanish sources.** Some languages render a Spanish text, not English: Belas is Spanish written in Greek letters (νιυ = de, ται = la), and several others carry a Spanish *Cloud of Unknowing*. - **Not word for word.** A few languages have more words than the English per sentence; positions do not line up. - **Sentence shape survives.** Sentence lengths and comma counts are kept almost exactly, so a post can be lined up with its source sentence by sentence even when few words are known. ## 4. Status, honestly About a third of the site is matched or readable; about a fifth of all words sit in sentences paired with their source. Most of the remaining posts are in small languages with no source yet. Absence of a source is not proof that a post is FL's own writing: many sources are simply not online. ## 5. Data files Base: https://codeberg.org/uncontrolledinfo/fl-sources/raw/branch/main/ | file | what it is | |---|---| | data/1_post_sources.csv | every matched post: post_id, FL link, language, date, title, source citation, how matched, confidence, sentences paired | | data/2_still_to_find.csv | sources known or suspected but not yet in hand, with what was ruled out | | data/3_identified_sources.csv | the bibliography (citation, DOI/arXiv/ISBN, posts covered) | | data/4_languages.csv | one row per language: posts, years, words known, readability, scripts, aliases/stages | | data/5_twins.csv | groups of posts that are the same text in different languages | | data/6_hunting_list.csv | well-decoded posts with no source yet, with their decoded opening | | data/7_word_lists.csv | all word lists in one table (large: ~70,000 rows) | | data/word_lists/.csv | one word list per language (spaces in names become _, e.g. Ladd_cryptolect.csv) | | data/reskin_examples.csv | original sentence vs FL version for confirmed reskins | | docs/methods.md | how matches and word lists were made and checked | The word lists include everything we use, each row labelled with its origin (how_learned): our own mining, the Forgotten Languages Discord group's lexicons (Purplexed1's Dirya list, spudd's Ned/Laari/Rhydd/Englabi lists, others unnamed), TrapDaddy's forgotten-languages-decoded glossaries (Yid, and Idhun -- most of what is readable in Idhun) and pilkingtonsbrain's lists ("Forgotten Languages, Decoded", incl. the Chinese sound table and Spanish lists). Citation: UncontrolledInfo, *Forgotten Languages: sources, word lists and methods*, 2026, https://codeberg.org/uncontrolledinfo/fl-sources (data CC BY 4.0).