# Departures from the pre-registration

Each is listed with its reason and with when it was made relative to the scoring of any
invented text. None was made after a tongue had been scored unless it says so.

## D1. Sanskrit is read in IAST (before any tongue was scored)

The corpus holds one Sanskrit New Testament in twenty-two scripts and romanizations. The
panel rule takes the first Latin-script candidate by translation id, which would be `sancol`
or `sanhk`: ASCII transliteration schemes that spell retroflex and aspirated consonants with
capitals and with spare letters. Lowercased, they would give Sanskrit a spelling no reader of
Sanskrit uses. The scholarly romanization, IAST (`sanias`), is taken instead, and the
normalizer then drops its diacritics like any other language's.

## D2. Which candidate is tested for script (before any tongue was scored)

The rule says a language enters the atlas if its New Testament text is at least 95% Latin
letters, one translation per language. The first build picked one translation per language
and then tested it, so a language whose first candidate was in another script was dropped
even when a later candidate was in Latin letters. The builder now passes over a candidate
that fails the test and tries the next in the same order (public domain first, then byte
order of the translation id).

## D3. Modifier letters are not counted in the Latin test (before any tongue was scored)

The first build dropped Tongan, a language written in Latin letters, at 91.2%: its glottal
stop is written with a modifier letter (Unicode category Lm), which is a letter and is not of
the Latin script. The test exists to keep out other scripts, and the normalizer folds such
marks to the apostrophe or drops them. Letters of category Lm are now left out of the count
on both sides of the ratio.

## D4. The replication corpus comes from a mirror (before any tongue was scored)

The pre-registration names unicode.org/udhr as the source of the Universal Declaration of
Human Rights in Unicode. On 2026-11-15 that address answered that the Unicode Consortium
stopped hosting the project in January 2024. The same collection is taken as NLTK
redistributes it (`corpora/udhr2.zip`), 388 text files. The mirror carries no table of
scripts, so the atlas's own test decides which texts are in Latin letters, and a text with
fewer than 200 types is left out.

## D5. The letter ratios are taken at matched totals (after the first tongues were scored)

M4 as registered adds one half to each letter count on both sides. A tongue has a few
thousand letters and a home lexicon has tens of thousands, so half a count is a far larger
share of the tongue: a letter that neither side uses at all came out as heavily over-used by
the tongue (w in Martian, with no w in it, at 44 times its French share). The flaw was seen
on the first run, in Martian's own top three. The home's counts are now scaled to the
tongue's total before the half is added, so a letter absent from both has a ratio of one.
Both versions are computed and kept (`m4registered`, `m4`); the predictions that rest on M4
(H4, H5, H11) are reported under both.

## D6. A second reading of "nearest", on each language's own ruler (after the first tongues were scored)

Control C3 was registered with the expectation that random letters would be attributed
nowhere in particular. They are attributed, and nearly always to the same few languages
(Vietnamese, Tok Pisin and English in the pocket panel, in 52 of the 54 noise controls of the
final run, and Haitian Creole in the other two; Bu, Sea Island Creole English and Naga Pidgin
in the atlas): the languages whose models are the most tolerant of anything. So
the raw nearest language (M1, kept as registered) can be the most tolerant ear instead of
the most similar tongue. An exploratory second reading is added beside it and labelled as
such everywhere: for every language L, the tongue's place on L's own ruler, that is, how
many of the atlas's real languages are nearer to L than the tongue is. `atlas-rulers.mjs`
computes each atlas language's ruler. No registered prediction is re-scored with it.

## D7. Two designed tongues are turned into letters by rule (Solresol after a first, wrong scoring)

Solresol was delivered as its grammar prints it, in a shorthand of single consonants ("frsod"),
and was first scored that way: 1.60 from French, farther than 718 of 729 real languages. That
measured the shorthand and not the language. The grammar gives the rule ("One can write just
d for do; r for re; m for mi; f for fa; so for sol; l for la; s for si"), and every form is now
expanded by it before scoring ("faresoldo"); the first figure is kept here and nowhere else.
Wilkins prints four vowel letters of his own, which the normalizer would treat as gaps; they
are read as the Latin letters they are shaped on (a, i, u, y), a choice fixed before his text
was scored. Both rules are in `tongues/designed/transliterate.mjs`. The same kind of step, a
fixed table, was already registered for the Russian tongues, which are romanized by the table
the Russian panel language is read with.

The Solresol rule was then corrected once more, before publication. The grammar writes a
capital for a doubled syllable ("To abbreviate these forms in writing, one uses capital D in
the place of dd, capital R for rr, M for mm, F for ff, SO for soso, L for ll, and S for ss"),
and the first version of the rule had read every capital as a single syllable. The author of
the tongue cards, an agent working from the kept sources, set the rule beside the tongue
file's own notes and reported the disagreement. A capital standing alone is now the doubled
tense word (D is dodo); a capital that only begins a longer form is ordinary capitalization,
as in the grammar's word lists ("Dore ou Dr, I, me"). Solresol went from 86 distinct words to
90 and from 0.57 to 0.58 from French.

## D8. Esperanto was registered and never gathered; tongues were added after the registration (found when the page was checked against the registration)

The registration's table of homes lists Esperanto (Zamenhof; Russian, then Polish). No text
of it was gathered: the brief for the designed languages left it out by the author's
oversight, and it was noticed only when an agent checked the page's sentences against the
registration, after everything was scored. No Esperanto tongue is scored. What the page
reports comes from the atlas, which turned out to hold Esperanto as one of its 730 languages:
the corpus has an Esperanto New Testament, and the rule that built the atlas did not ask
whether a language had been constructed. The page gives the place of that translation's
held-out words on the Russian and Polish rulers, by the method used for every real language
on another's ruler (METHOD.md, section 12). That is a translation by other hands, not
Zamenhof's own writing, and it was never registered as a tongue; it is labelled as what it is,
and it enters no prediction. The atlas was left as registered, Esperanto included, and the
page says that one of its real languages is an invented one.

The registration fixed a home for each maker then on its list, and "as recorded" for the
people recorded speaking in tongues. Tongues found afterwards were given homes in
`tongues/registry.json` by the registration's rule ("the language the maker spoke or wrote")
before they were scored: English for Wilkins, who wrote in it; English for Lojban's root
words, the language of the documents that define them (a guess of the same kind as the
models' home); Russian for Kruchenykh; French for Hélène Smith's other tongues. For Pfister's
three cases the article states no mother tongue (it explains words of Simon's and Lina's from
Swiss German), and for M'Kerrell's speaker the printing states none: they are scored against
German and English, the languages of the texts that report them. The 1912 respelling of the
Enochian Calls is scored beside the 1659 printing and counted in no prediction.
