The old v2 engine was useful for diagnostics, but its rule system was too shallow. It behaved like a decorative regex layer instead of 20,000 years of language history.
The v3 model treats evolution as four 5,000-year epochs. Each epoch has a different balance of pressure:
- Early Divergence: regular inherited sound change from Romance sources.
- Expansion Contact: pluralistic spacefaring contact, loanword pressure, local drift.
- Imperial Standardization: Boardroom/legal normalization and institutional compression.
- Liturgical/Trade Split: Temple fossilization, Boardroom polish, Trade erosion.
Every generated word must remain traceable to a real source, but the final form should not look like a lightly edited Romance word.
The engine should combine:
- deterministic sound rules
- seeded organic variation
- cultural pressure profiles
- loanword strata
- register-specific standardization
- diagnostics for weak divergence
Randomness must be reproducible. The same source, meaning, register, and seed should produce the same form.
This gives us organic irregularity without authorial hand-waving.
Each register pulls the language differently.
Temple Common:
- preserves archaisms
- restores vowels for rhythm
- allows long vowels and nasal vowels
- prefers solemn cadence
Boardroom Common:
- compresses but preserves precision
- favors hard consonants and stable legal forms
- standardizes competing variants
Trade Common:
- erodes aggressively
- simplifies clusters
- accepts loans readily
- optimizes for fast cross-world speech
Loanwords should not be random decoration. They should enter through domains:
- navigation
- trade
- law
- military
- temple
- administration
- frontier life
The app should eventually support domain-tagged donor strata.
A candidate should fail if:
- too few transformations happen
- source similarity remains too high
- the final form is too long and transparent
- the form violates register expectations
manana should not become manan after 20,000 years. That is a failure state.