conlang

A partial-spec constructed-language generator.

You describe only what you care about — phonology, syllable shapes, a handful of words — and the generator samples the rest from typologically-plausible defaults. Same spec plus same seed always produces byte-identical output, and sampled entries can be locked so they survive re-sampling.

spec (YAML) ──▶ generate ──▶ model (canonical JSON) ──▶ render ──▶ dictionary + grammar sketch
                                   │
                                   ├── lint   (machine-checkable invariants)
                                   └── bench  (plausibility scorecard vs. Kah)

The model JSON is the single source of truth; the dictionary, grammar sketch, and any future views are rendered from it.

Install

pip install conlanggen

The distribution is conlanggen, but the import package and the command are both conlang — import conlang, conlang generate.

Quickstart

python -m venv .venv
.\.venv\Scripts\python.exe -m pip install -e ".[dev]"

# Generate from the example spec, and render a Markdown dictionary.
.\.venv\Scripts\python.exe -m conlang.cli generate --spec spec\example.yaml --render

# Happy with it? Freeze it into an explicit spec so the values stick.
conlang accept --model out\veshtari.json

Output lands in out/: a model JSON and a rendered .md.

# Re-lint or re-render an existing model.
conlang lint --model out\veshtari.json
conlang render --model out\veshtari.json --out out\veshtari.md

Translating a fixed story

The toolkit ships a repeatable pipeline for translating the fixed 111-sentence story The Quiet Morning into a generated language: conlang new → grow → translate → validate-translation → extras. The runbook is docs/pipeline.md.

Writing a spec

Everything is optional. Omitted fields are sampled.

name: Veshtari
seed: 42

phonology:
  consonants: [p, t, k, m, n, s, l, r, v, sh, ch]
  vowels: [a, e, i, o, u]

syllables: [CV, CVC, V]   # C = consonant slot, V = vowel slot
word_count: 150

lexicon:                  # pin specific words; they are locked by default
  - gloss: water
    root: vela

Accepted keys: name, seed, phonology.{consonants,vowels,ipa,orthography}, syllables, concepts, word_count, lexicon, affixes, onsets, codas, forbidden.

An explicit, unsized lexicon is treated as the complete word list (that is what accept writes back); otherwise the built-in concept list is used and word_count trims or extends it.

If your syllables include CCV/CCVC, the generator samples a legal onset set unless you supply onsets yourself, and accept records whichever was used so a frozen spec regenerates identically.

IPA inventories

By default a symbol is its spelling (sh is written sh). Set phonology.ipa: true to declare the inventory phonemic: each symbol is an IPA phoneme, and the tool derives a romanised orthography for the dictionary.

name: Shibo
seed: 11

phonology:
  ipa: true
  consonants: [p, t, k, ʃ, tʃ, m, n, s, l]
  vowels: [a, e, i, o, u]
  orthography:        # optional; corrects or extends the built-in defaults
    tʃ: ch
    ʃ: sh

/ʃ/ spells sh and /tʃ/ spells ch from a built-in table; anything absent spells itself, so a plain Latin inventory is unchanged. orthography overrides the table (set an entry back to its own symbol to cancel a default), and the resolved map is stored on the inventory. The dictionary and word-form tables then show a spelling column beside the ipa one, and accept writes the map back so a frozen spec regenerates identically. Lint warns on a map entry that is not an inventory symbol, or two symbols that would spell the same thing.

Freezing a language (accept)

A partial spec samples whatever you leave out. Once you like the result, conlang accept writes it back as a fully-explicit spec: the phonology, syllable shapes, constraints, affixes, and every root become pinned entries. Regenerating from that spec reproduces the same language exactly — sampled values now survive re-sampling, and changing the seed can no longer disturb them.

conlang generate --spec spec\example.yaml --out out\veshtari.json
conlang accept  --model out\veshtari.json --out spec/veshtari.yaml
conlang generate --spec spec/veshtari.yaml --out out/veshtari.json

Two fields legitimately change across the round trip: origin becomes specified for every accepted entry (it is pinned now, not sampled), and spec_hash changes because the spec text is different. Everything linguistic is identical, and tests/test_invariants.py asserts exactly that.

Guarantees and lint rules

Reproducibility is the one hard guarantee: spec + seed ⇒ identical bytes. Everything else is reported by lint as errors or warnings, and a broken language is still generated (conlangers want to explore broken states).

Every rule above has a test in tests/test_lint.py, tests/test_lint_invariants.py, or tests/test_morphology.py.

Phonotactics

A word is legal when its C/V skeleton can be partitioned into the language's syllable templates and every resulting onset and coda is allowed. Two details matter more than they look:

• Constraints are checked while syllabifying, not after. A word often has several valid partitions; deplango is de|plan|go with legal onsets, and also de|pla|ngo with an illegal one. Verdict must depend on the phonology, not on which partition the search happens to return first.
• Onsets and codas are compared by segment, not by string. An onset of ny + m renders as "nym" — three characters, two segments. Matching strings would silently reject legal words in any language with a digraph.

When a template set admits onset clusters (CCV, CCVC), the generator samples a legal onset set and builds roots from it, so generated words are legal by construction rather than by luck. constraints.onsets in the model records what was sampled; an explicit onsets in the spec always wins.

Affixes carry allomorphs: conditioned surface forms. A single form cannot serve both consonant-final and vowel-final roots, because the root's coda plus a consonant-initial affix resyllabify into an onset cluster. When the primary form would be illegal, build_words uses the first alternative that is legal — which is what real languages do. It is a rare fallback (~2 in 8,500 affixed forms), but without it those words would be illegal and lint would fail the language.

The Kah benchmark

Kah (Kwesho's language) is hand-made, so the benchmark does not try to regenerate it. It does two things:

1. Adapter fixture — parses the real Kwesho dictionary cache into the model schema, proving the schema can represent real data.
2. Plausibility scorecard — compares a generated language against Kah.

The scorecard deliberately reports three separate numbers, because a word can fail for two unrelated reasons and collapsing them makes the report unreadable:

Scoring the built-in kah preset against Kah is circular, so the report says so explicitly when it detects it. Treat the scorecard as a lever, not a target: the presets are deliberately small, and a language that only allows CV syllables is a legitimate (if monotonous) language type, not a bug to tune away.

conlang bench --spec spec\example.yaml
# or point at a cache explicitly:
conlang bench --spec spec\example.yaml --kah C:\path\to\kah-lexicon.txt

Kah's dictionary is Kwesho's copyrighted work and is never vendored into this repository. Without --kah, the benchmark looks for the glossary at a per-user cache path (%LOCALAPPDATA%\conlang\kah-lexicon.txt on Windows, ~/.cache/conlang/kah-lexicon.txt elsewhere, or $XDG_CACHE_HOME/conlang/); override it with --kah or CONLANG_KAH_LEXICON.

On the reference machine, the adapter parses 8,153 entries with a 97.8% validity rate.

Evolving a language

A generated model can be the parent of a daughter, evolved by ordered sound changes. A rule is old > new (new empty deletes the segment), optionally conditioned by an environment after /. The environment names the context on each side of the target, with _ marking the target slot, # the word boundary, and V/C the vowel/consonant classes:

conlang generate --spec spec\example.yaml --out out\proto.json
conlang evolve --model out\proto.json --rule "a > e" --rule "p > f / V_V" --out out\daughter.json
conlang evolve --model out\proto.json --rule "t > d / #_" --rule "s > z / V_" --out out\voiced.json

V_V is "between vowels", #_ is word-initial, _# is word-final. Rules are applied left to right so a later rule sees an earlier one's output (feeding).

The daughter is a full model: its inventory gains any new sounds, its lexicon and affixes are rewritten, and its word forms are rebuilt. It records where it came from in lineage:

"lineage": { "parent": "proto", "rules": ["a > e", "p > f / V_V"] }

That is the whole point of the field: the derivation is data, so a daughter can itself be evolved, and accept keeps the chain across a round trip. Mergers introduced by a change show up as lexicon.homophones warnings, never errors.

Syntax

Each language samples a constituent order (SOV, SVO, or VSO, weighted toward the first two) and an alignment (accusative or ergative). Both are recorded in the model and survive accept. A transitive clause is linearised from the lexicon and inflected: the subject and object take case by alignment (accusative marks NOM/ACC; ergative marks ERG/ABS) and the verb agrees with the subject:

from conlang.syntax import sample_clause
words, gloss = sample_clause(language)
# e.g. (['pyople', 'klavragprekbra', 'blonyto'], 'water.ERG say.3SG fire.ABS')

Case and agreement are applied when the clause is built, not woven into every word's standing paradigm — a deliberate split, so a noun's form count does not double again for case.

Layout

src/conlang/
  model/        canonical pydantic schema (the source of truth)
  sampling/     deterministic RNG, inventory presets, concepts, defaults
  phonology/    tokenization, syllabification, legality checks
  lexicon/      seeded root generation, stable per-item seeding
  morphology/   affixes, inflection classes, and word-form assembly
  soundchange/  ordered sound changes; derive a daughter language
  syntax/       word order, alignment, simple clause linearisation
  render/       Markdown dictionary + grammar sketch
  lint/         machine-checkable invariants
  bench/        Kah adapter + scorecard
  cli.py        thin command-line view over the library

Development

.\.venv\Scripts\python.exe -m pytest
.\.venv\Scripts\python.exe -m ruff check .
.\.venv\Scripts\python.exe -m mypy src

Roadmap

Shipped so far:

• Phonology — inventory, phonotactics, seeded word generation.
• IPA inventories — declare the inventory phonemic (ipa: true) and the tool derives a romanised orthography alongside the phonemic forms.
• Morphology — affix breadth, inflection classes, and stacking (a word can carry several categories at once).
• Diachrony — conlang evolve: ordered sound changes derive a daughter language, now with environments (p > f / V_V, #, _), and the chain recorded in lineage.
• Syntax — sampled word order and alignment, case marking with subject agreement, the predication tier (adpositions, noun phrases, intransitives, copula/existential), clause sequencing (parataxis and conjunctions), and aspect (progressive/perfect, stacking with tense).
• Translation pipeline — conlang new / grow / translate / validate-translation / extras: build a generated language and translate the fixed story into it, including a grammar-match port from a sibling language.
• Writing systems — conlang script: a language samples a script (alphabet / syllabary / impure abjad / featural) from its phonology, the tool generates deterministic SVG glyphs as a derived view, and renders a script chart plus the lexicon written out (the recall deck shows glyphs too).
• Data packs — optional, pinned PHOIBLE (inventories + features) and WALS (typology) ground sampling in real data (conlang data fetch), with a graceful fallback to the built-in presets when absent.
• Feature-driven phonology — natural classes (obstruent, nasal, …) usable in onset/coda constraints, and feature-based sound-change rules ([-sonorant, -continuant] > [+continuant] / V_V) in conlang evolve; a changed vector resolves back to a symbol via the feature layer.
• Place structure — each predicate carries an ordered list of semantic roles (Lexeme.places): verbs 1–3 places, the copula THEME, ATTRIBUTE, adpositions THEME + a spatial role. A ## Place structure view lists them, and clauses are role-labelled (water.ERG.AGENT give.3SG fire.ABS.PATIENT) — orthogonal to case.
• Semantics — the logical operators (negation as a preverbal particle, quantifiers as NP modifiers, coordination with or/nor) and information structure (topic / focus markers), all closed function-word classes that add a word without reordering anything; a ## Semantics view lists the inventories and worked clauses.

Next:

• Formal semantics (truth-conditional model theory) — deferred.

License

MIT (code). Kah dictionary and grammar content remains Kwesho's and is not part of this repository.