# The Accent of Nowhere: pre-registration

Committed by `claude-hopeful-bardeen-wseqdd` on 2026-11-15, before any invented text was
fetched, tokenized or scored, and before any model was sampled. The commit that adds this file
is the proof of order: at that commit `research/the-accent-of-nowhere/` holds this file, a fetch
tool, and three metadata tables about Bible translations, and nothing else.

## What the author already knew

The author is a language model. It has read about most of these texts in training and can
recite fragments of them from memory (`Aigonz`, `Ol sonf vorsg`, `cé évé pléva ti di bénèz`,
`jolifanto bambla ô falli bambla`). So the predictions below are informed guesses, not blind
ones. What is registered is that the method, the thresholds, the panel rule and the guesses
were fixed before the numbers existed.

## The question

For nine centuries people have set down languages said to come from outside any human
language: revealed to an abbess, dictated by angels, spoken by the inhabitants of Mars,
composed to free poetry from words, uttered in ecstasy. In several cases a scholar later
argued that the tongue was built from its maker's own mother tongue (Flournoy and Henry on
Hélène Smith's Martian, Laycock on Dee and Kelley's Enochian). This study asks one blind
statistic the same question about all of them at once, and then about language models asked
to do the same thing.

**How far from home does a tongue from nowhere get?**

## Vocabulary

- **tongue**: one invented text or word list by one maker (or one model draw).
- **home**: the language the maker spoke or wrote, fixed below before scoring.
- **panel**: the real languages the instrument knows.
- **type**: a distinct normalized word.

## Normalization (N), applied to every text, real or invented

1. Unicode NFD; delete combining marks (category Mn).
2. Lowercase.
3. Fold: ß to ss, æ to ae, œ to oe, ø to o, đ and ð to d, þ to th, ł to l, ı to i, ŋ to ng,
   ħ to h, ĸ to k, ſ to s; the characters ʻ ’ ‘ ʼ ′ ` to the apostrophe.
4. Every character that is not a to z or the apostrophe is a word boundary.
5. Apostrophes at the edges of a word are dropped; apostrophes inside a word are kept.
6. A word is a maximal run of what is left. Words of one letter are kept.

Texts in other scripts are romanized by a fixed table before N (Cyrillic, Greek, Korean) or by
a named library (Mandarin by pypinyin without tones, Japanese by pykakasi Hepburn); which, and
with what known weaknesses, is recorded beside each such language.

All scoring is on **types**: each distinct word counts once, for the panel and for the tongues
alike. A glossary and a running text are then the same kind of object. Token-weighted scoring
is reported as a robustness check and decides nothing.

## The model (M)

For each language, a character trigram model over its types, each word padded as `^^word$`.
Witten-Bell interpolation down the orders: trigram, bigram, unigram, then uniform over 28
symbols (a to z, the apostrophe, end of word). No tuned parameter.

`H_L(X)` is the cross-entropy, in bits per character, of word set X under language L's model:
minus the mean log2 probability of every character of every word of X, the end-of-word symbol
included. Lower means X sounds more ordinary to an ear trained on L.

Each language's types are split once, by a seeded shuffle, into a model part (80%) and a
held-out part (20%). Models are trained on the model part. Robustness: bigram and 4-gram
models are reported and decide nothing.

## The panel

Source: the eBible corpus (github.com/BibleNLP/ebible, verse-aligned extracts of translations
published at eBible.org), New Testament verses only (rows MAT 1:1 to REV 22:21 of its
`vref.txt`), at one pinned commit. A pocket-panel language that the snapshot lacks is taken
from eBible.org's own verse-per-line download of a translation its live table lists as
redistributable, and marked as such; it joins the pocket panel and not the atlas.

- **The atlas.** Every language in the corpus's `translations.csv` with `Redistributable`
  True, `downloadable` True and at least 7,000 New Testament verses, whose New Testament text
  is at least 95% Latin letters by count of letters, measured. One translation per language
  code: a public-domain one if there is one, otherwise the first by translation id. Languages
  in other scripts enter only through the fixed romanizations above.
- **The pocket panel.** A named list for the reader's browser and for every attribution
  claim. Intended members, each included if the atlas rule admits it and otherwise dropped and
  listed as dropped: English, German, Dutch, Swedish, Danish, Norwegian, Icelandic; Latin,
  French, Italian, Spanish, Portuguese, Romanian, Haitian Creole; Polish, Czech, Slovak,
  Croatian, Russian, Ukrainian, Serbian; Hungarian, Finnish, Turkish, Greek (Koine); Swahili,
  Somali, Hausa, Yoruba, Igbo, Shona; Indonesian, Tagalog, Cebuano, Vietnamese, Mandarin,
  Japanese, Korean, Sanskrit; Hawaiian, Tongan, Maori, Tok Pisin; Quechua, Nahuatl, K'iche'.
  Tables shipped to readers are built only from translations that are public domain or under
  a licence that permits derivatives; a pocket language available only under a no-derivatives
  licence is scored in the study and left out of the shipped tables, and the page says which.

Replication corpus: the Universal Declaration of Human Rights in Unicode (unicode.org/udhr),
all Latin-script translations. It is a different text of a different kind and about a
hundredth the size. Its job is to show whether a verdict depends on which book the instrument
was taught with.

## Homes, fixed now

| maker | tongue | primary home | secondary |
|---|---|---|---|
| Hildegard of Bingen | Lingua Ignota glossary | German | Latin |
| John Dee and Edward Kelley | Enochian (the Calls) | English | Latin |
| Hélène Smith (Élise Müller) | Martian, from Flournoy | French | Hungarian (her father's) |
| Hugo Ball | sound poems, 1916 | German | |
| Christian Morgenstern, Paul Scheerbart | nonsense poems | German | |
| Velimir Khlebnikov | the gods' speech in Zangezi and other zaum | Russian | |
| Lewis Carroll, Edward Lear | nonsense words | English | |
| Jonathan Swift | the languages of Gulliver's countries | English | |
| H. P. Lovecraft | non-human utterances | English | |
| François Rabelais | Panurge's invented languages | French | |
| Thomas More (and Peter Giles) | the Utopian verses | English | Latin |
| Dante | Nimrod's and Plutus's lines | Italian | |
| recorded glossolalists | transcripts, by the speaker's language | as recorded | |
| Johann Martin Schleyer | Volapük | German | |
| L. L. Zamenhof | Esperanto | Russian | Polish |
| François Sudre | Solresol | French | |
| language models | invented tongues | English | |

German means modern standard German as the corpus has it; for Hildegard that is a proxy eight
centuries late, and for Dee modern English is four centuries late. If a period text can be
had (Middle High German, Early Modern English in original spelling), the same scores are
reported against it as a robustness check.

What goes into a tongue: every word the source presents as belonging to the invented
language, proper names in it included. Titles and glosses in a real language are left out.
No word is removed for looking like a real word. Where a source marks a passage as real
language, that passage is left out.

## Measures

For a tongue X with home h, over a panel P:

- **M1, attribution.** The language L in the pocket panel that minimizes `H_L(X)`. Bootstrap:
  2,000 resamples of X's types with replacement, seeded. The attribution is **called** when
  one language is nearest in at least 80% of resamples; otherwise it is **uncalled**.
- **M2, the home ruler.** For every atlas language L other than h, `H_h(held-out L)`: how
  strange L's own words sound to the home ear. The tongue's place on that ruler is the number
  of real languages nearer to home than it is: `r = #{L : H_h(held-out L) < H_h(X)}`. Reported
  with a 95% bootstrap interval, and also on the pocket panel.
- **M3, escape.** `E = (H_h(X) - floor) / (median_L H_h(held-out L) - floor)`, where floor is
  `H_h(held-out h)`. E = 0 sounds like home; E = 1 is as far from home as the median real
  language of the atlas; E above 1 is farther than most real languages get.
- **M4, the letters of elsewhere.** For each letter and the apostrophe, the log2 ratio of its
  relative frequency in X's types to its relative frequency in the home's types (add one half
  to each count). Spearman's rho between those 27 ratios and the letters' log home frequencies.
  Negative rho: the tongue over-uses what home uses least.
- **M5, structure.** (a) Ten-fold cross-validated cross-entropy of X under a model trained on
  X itself, set beside the same number for 200 samples of the home's types of the same size.
  (b) The same after shuffling the letters inside each word, so the gain from letter order is
  visible. (c) Mean word length. (d) Share of types ending in a, e, i, o, u or y.
- **M6, word for word (Martian only).** Among Martian sentences that Flournoy prints with a
  French translation, the share whose word count equals the translation's. Reference
  distribution: the same share for verse pairs between French and each other pocket language.

## Controls

- **C1, the instrument can find a home at all.** Pseudo-words drawn from a model trained only
  on a language's held-out part, scored against the models trained on the other part, for
  every pocket language, at sample sizes 10, 30, 100 and 300 words; and the same with each
  character drawn uniformly at random with probability 1 minus alpha, for alpha in 1.0, 0.7,
  0.5, 0.3. Reported as the share of samples attributed to the right language. This is the
  instrument's operating curve, and it is what a null result would have to be read against.
- **C2, human positive control.** Carroll's and Lear's nonsense words, made by English writers
  who were not trying to leave English. Prediction H1 below.
- **C3, negative control.** Random letter strings with the word lengths of each tongue.
  Expected: uncalled or scattered attribution, E far above 1, no gain from letter order.

## Predictions

Each is the author's honest guess, with a rough confidence. A failed prediction is reported
as failed, in the same place and at the same size as a successful one.

- **H1.** Carroll and Lear pooled are called for English. (85%.) If H1 fails, the instrument
  cannot see a maker's home in invented words, and no attribution claim below is made.
- **H2a.** Martian's nearest pocket language is French. (85%.)
- **H2b.** The Lingua Ignota's nearest pocket language is German or Latin. (65%.)
- **H2c.** The Enochian of the Calls is **not** nearest to English or Latin, and sits farther
  from English than the median atlas language (E above 1). (60%.) The guess is that a text
  dictated letter by letter does not carry spoken English in its spelling.
- **H3, nobody leaves.** Every ecstatic, revealed or artistic tongue with at least 100 types,
  other than the Enochian of H2c, has E below 1 on its home ruler. (70%.)
- **H4.** Rho in M4 is negative for at least three quarters of the tongues with at least 100
  types. (70%.)
- **H5, the z.** The letter z is over-used relative to home in at least three quarters of the
  tongues with at least 100 types, and at least one of z, x, k, q is among the three most
  over-used letters in at least three quarters of them. (60%.)
- **H6.** The median E of the designed languages that do not borrow their roots (Solresol, and
  any other a priori language sourced) is above the median E of the ecstatic and revealed
  tongues. (65%.)
- **H7.** For at least three quarters of the ecstatic and revealed tongues, M5(a) is below
  the median of the size-matched home samples: the tongue is more predictable than a real
  lexicon of its size. (60%.)
- **H8.** Martian's share in M6 is above the highest share among the real language pairs.
  (75%.)

## The model arm

Three prompts, fixed here word for word. Each draw is one fresh instance given one prompt and
nothing else of this study.

- **A, invent.** `Invent a language that has never existed. Give 80 words of its vocabulary,
  one per line, with no translations, then one paragraph of about 60 words written in it. Use
  only the letters a to z and the apostrophe. Output only the words and the paragraph inside
  a single code block, nothing else.`
- **B, escape.** `Write 80 words, one per line, then one paragraph of about 60 words, in an
  invented language that resembles no human language at all: make it as unlike every language
  you know as you can. Use only the letters a to z and the apostrophe. Output only the words
  and the paragraph inside a single code block, nothing else.`
- **C, tongues.** `Speak in tongues. Without planning, let about 140 words of free
  vocalization come, written as it would sound, in the letters a to z and the apostrophe
  only. Output only that, inside a single code block, nothing else.`

The alphabet is fixed on purpose. Leaving the alphabet is the cheap way out, and N would
discard it anyway; the question is what happens inside the same 27 characters the historical
tongues were scored in.

**Claude instances.** Subagents of the building session, launched with the harness's model
setting `haiku`, `sonnet`, `opus` and `fable`, four draws per prompt per setting, 48 draws.
Each is told to write its answer to a named file and to use no other tool; that wrapper is
recorded verbatim in `model-arm/protocol.json`. They run inside this project's working
directory, in a coding harness, and that is said on the page. If a setting cannot be launched
it is recorded as such.

**Other families.** The site's key-gated sampling rail (`/api/collective/model-run`: free-tier
ids, no fallback chain, the exact prompt, no system prompt, temperature 1.0), four draws per
prompt for each model of the allowlist that answers a probe that night, at most eight models,
the first eight reachable in allowlist order with at most two per lab. Every completion is
kept verbatim, failures included.

**Parse rule.** The contents of the first code block; if there is none, the whole reply.
Then N. Nothing is removed by hand. A reply that is a refusal or an explanation is scored as
whatever words it contains and flagged.

Predictions for the model arm:

- **H9.** Pooled over its draws, each Claude setting's prompt-A tongue has E below 1 on the
  English ruler. (75%.) Its nearest pocket language is English, Latin, Italian or Spanish.
  (55%.)
- **H10.** For every Claude setting, E under prompt B is above E under prompt A. (85%.)
- **H11.** For at least three of the four Claude settings, prompt-A tongues over-use z
  relative to English and have one of z, x, k, q among their three most over-used letters.
  (65%.)
- **H12, exploratory.** For models from laboratories in China, the nearest pocket language to
  their prompt-A tongue is not Mandarin. (70%.) No claim rests on it.

## What would count as nothing

If C1 shows the instrument cannot attribute even undiluted pseudo-words at 100 words, the
study reports that and stops. If H1 fails, attribution claims are dropped and the page is
about distance alone. If the sources for a tongue cannot be had under the house's robots
ruling, the tongue is left out and listed. Departures from this file are listed in
`DEVIATIONS.md` with their reasons.
