How we build languages

Record 01

Each voice starts as a structured profile

Every voice is defined in its own TypeScript record. That record names the voice and stores its vocabulary, exact phrase matches, grammar settings, sound substitutions, presentation style, unknown-word strategy, examples, and public disclaimer. The shared engine reads the same shape for every voice, so differences live in reviewable data rather than hidden prompts.

What each current voice is designed to do

These briefs come from the same public tool definitions used by the language index. The fallback column is read from each matching language profile.

Record 02

The translation path, in order

1. 01

   Normalize the input

   Trim whitespace, collapse repeated spaces, remove HTML-like tags, and keep at most 500 characters.

2. 02

   Look for a phrasebook match

   An exact normalized phrase takes priority and returns the wording stored in that profile directly.

3. 03

   Tokenize and apply grammar

   Words, spaces, numbers, and punctuation are separated. Active grammar rules run while the words are still English.

4. 04

   Translate or transform each word

   After guards for probable names, existing target vocabulary, and Robot function words, known words use the dictionary. Others follow the profile’s fallback strategy.

5. 05

   Apply the voice’s finish

   Capitalization, punctuation, optional framing, and stable flavor words create the final cadence before spacing is tidied.

What happens to an unknown word?

It is not silently deleted. Most voices rebuild it with their configured letter substitutions and stable affixes. Robot labels it as a unit; Medieval checks a small modern-to-period map before adding a stable ending; fantasy-root profiles construct a compact invented root; and a preserve strategy can leave it unchanged. The same input and voice produce the same result.

A capitalized word in the middle of a sentence is treated as a possible proper noun and restored after style changes. This is a practical heuristic, not full name recognition.

Record 03

Examples generated by the current engine

These original sample lines are rendered by the same server-side function used by the product. They are not manually polished for this page, so they remain an honest snapshot of current behavior.

Record 04

Where accuracy stops

• These are invented voices. Output is not an official, canonical, academic, or certified translation.
• Unknown-word transformations imitate form, not meaning.A word can sound consistent with a voice without carrying a real semantic translation.
• Grammar is deliberately limited. The SOV rule is a short-sentence heuristic that assumes the second word is a verb; there is no full grammatical parser.
• Confidence is an internal coverage label. It is calculated from phrasebook or dictionary coverage and must not be read as a percentage of linguistic correctness.
• English is the working source language. The dictionaries and grammar filters are written around English input.

Record 05

How review and corrections work

A suggestion is never used to claim an official spelling. We assess whether it improves the intended fictional voice, remains internally consistent, and avoids protected text or implied affiliation. A change then belongs in the versioned profile or shared engine, where its effect on other words and phrases can be checked.

The repository’s automated checks verify deterministic, non-empty output across every voice, output for the selected popular phrases, provider routing, public response boundaries, shared-message behavior, moderation, route generation, and the production build. They catch technical regressions; they do not replace editorial judgment about clarity, originality, or usefulness.