# Vumarii 2019: Defango-listed Esperanto / Swahili controls

## Scope

The preserved 2019 community `Solver key 1.0` visibly lists `Esperanto` and `swahili` near Caesar, Vigenere and Atbash. It is not an author-signed Vumarii key. This audit therefore treats the labels as two narrowly stated community hypotheses, not as a discovered plaintext or as evidence that the author selected either language.

For each language, Tatoeba's public language-tagged sentence export is archived byte-for-byte, split deterministically by sentence ID, and used to train a held-out A-Z character 5-gram model. The text accepts the red image legend and the seven community readings only conditionally, then exactly enumerates the 8P5 = 6,720 injective assignments of the five remaining glyph classes.

## Result

### Esperanto

- Corpus: `818623` normalized sentences; `21393606` training and `5353125` held-out letters; source snapshot SHA-256 `2bb2ad216e22c9fda0086b784915d312e2061e5dd6cac348fbd9e5acd25d6459`.
- Best of `6720` completions: C16/C17/C18/C20/C21 = `W S H K Z`; `-6.1601` bits/character; held-out z = `-18.05`; percentile = `0.00`.
- Position-shuffle control: `0` / `256` maxima at least as high (empirical p = `0.0039`).
- Seven-reading selection control: `0` / `256` maxima at least as high (empirical p = `0.0039`).
- Best displayed candidate: `TVI ULTO DA VUMARII DO CERPO TA TENGRI IGWA GENENNE NSOP HODANE DE QUINNA LANG KLANA CAUZU ABILTO DAUZ CSAKAT ANDAMAK TENGRI ITKSIN MEP GENENNE TLSOK KAPIHT ETMEK DE CSED WEZARU SEDOFER`.

### Swahili

- Corpus: `4583` normalized sentences; `149421` training and `37124` held-out letters; source snapshot SHA-256 `64a16e513c5277772827d11a89fedb516a2a88df560690817ee95a82480ea70a`.
- Best of `6720` completions: C16/C17/C18/C20/C21 = `W K S H J`; `-6.0662` bits/character; held-out z = `-13.91`; percentile = `0.00`.
- Position-shuffle control: `0` / `256` maxima at least as high (empirical p = `0.0039`).
- Seven-reading selection control: `13` / `256` maxima at least as high (empirical p = `0.0545`).
- Best displayed candidate: `TVI ULTO DA VUMARII DO CERPO TA TENGRI IGWA GENENNE NKOP SODANE DE QUINNA LANG HLANA CAUJU ABILTO DAUJ CKAHAT ANDAMAH TENGRI ITHKIN MEP GENENNE TLKOH HAPIST ETMEH DE CKED WEJARU KEDOFER`.

## Interpretation boundary

A result that falls near held-out prose would only make a particular conditional surface worth follow-up; it still would not prove a translation. A result far below held-out prose, or one explainable by secondary-reading selection, rejects the corresponding simple monoalphabetic completion as readable prose under this model. Neither outcome tests arbitrary transliteration, a polyalphabetic/keyed cipher, an unknown language, or the upper diagram.

The author-selected evidence still consists of the image and red legend. The seven additional glyph readings and these two language names come from a community solver artifact. No source-selected key, plaintext, or translation is present in the archive.

## Reproduce

```powershell
& '<bundled-python>' scripts\audit_vumarii_2019_defango_language_hypotheses.py --fetch
```

Word-class source SHA-256: `81262c1ca605c686da3acba6b4ac1ea2f11196cf74e530a9c71f607d0c5c76f8`
Solver-key image SHA-256: `e5f59aa67ef674364e87969a8ab1fd35b873e26d750643748b79dcf075b28e9d`