Reference: Three generation levels

The invariant: a Spec is partial by design. Anything you omit is sampled. The same Spec plus seed produces the same model bytes. See the Spec reference.

The three levels

Level	What the Spec pins	What is sampled	Why choose it
1 · Minimal `level1-minimal`	`name`, `seed`	Inventory, templates, constraints, word count, lexicon, morphology, syntax, discourse, semantics	Fast exploration. Useful when you want a surprise or a starting point, not a designed language.
2 · Partial `level2-partial`	Inventory, syllable templates, word count	Lexicon, morphology, syntax, alignment, discourse, semantics	Control the sound and size while leaving the vocabulary and grammar to the generator.
3 · Rich `level3-rich`	IPA inventory, templates, forbidden cluster, concepts, lexicon roots, syntax, discourse, semantic model	Most affix forms, inflection classes, and the remaining word forms	Designed output: a specific sound, grammar, and interpretation with little sampling left.

What the three runs produced

Signal	Minimal	Partial	Rich
Seed	101	202	303
Lexemes	249	139	101
Word forms	1,466	584	184
Inventory	Large sampled IPA-like set	Ten Latin consonants, five vowels	Nine IPA consonants, five vowels
Orthography	Identity	Identity	IPA → spelling map for `ʔ` and `ʃ`
Word order	SOV	SOV	SOV
Alignment	Accusative	Ergative	Ergative
Modifier order	Prenominal (derived)	Prenominal (derived)	Postnominal (explicit)
Affixes	18	18	20
Discourse features	Off	Off	Particles, switch-reference, reflexives, evidentiality
Semantic model	None	None	Two individuals; `sun`, `moon`, `see`
Lint	Valid	Valid	Valid

Lexicon size is not the richness axis. Level 1 is the largest because it uses the default word count and full concept list. Level 3 is smaller because its concept list was pinned to twelve items. The axis is how much is specified, not how many words result.

The steps, once per language

# 1. Write a Spec at the level you want.
#    Only spec_version is structurally required.

# 2. Generate. --strict makes lint errors exit non-zero.
.\.venv\Scripts\python.exe -m conlang.cli generate \
  --spec out\language-levels\spec\level-3-rich.yaml \
  --out out\language-levels\level-3-rich.json \
  --workspace out\language-levels --strict

# 3. Render the grammar sketch.
.\.venv\Scripts\python.exe -m conlang.cli render \
  --model out\language-levels\level-3-rich.json \
  --out out\language-levels\level-3-rich.md \
  --workspace out\language-levels

# 4. Freeze it later if the sampled values are the ones you want.
.\.venv\Scripts\python.exe -m conlang.cli accept \
  --model out\language-levels\level-3-rich.json

What each kind of pinning buys you

Pinned field	Effect on the model
`phonology`	Controls the segment inventory and whether spellings are derived from IPA.
`syllables`	Controls legal word shapes and the templates used for roots and affixes.
`word_count`	Trims or extends the built-in concept list when no explicit lexicon is given.
`concepts`	Replaces the built-in concept list and fixes the lexical field.
`lexicon`	Pins roots, parts of speech, and lock status for chosen words.
`syntax`	Sets word order, alignment, modifier side, and subordinator edge.
`discourse`	Turns on particles, switch-reference, reflexives, and evidentiality.
`semantics`	Declares a domain and predicate extensions for evaluation and entailment.

Provenance: did the pin stick?

Model evidence	Meaning
`origin: specified`	The value came from the Spec.
`origin: sampled`	The generator chose the value.
`locked: true`	The entry should survive regeneration.
`concepts`	Selects which words exist; roots are still sampled.
`lexicon`	Pins a gloss, root, and part of speech.

In level-3-rich.json, lex:sun is specified and locked; lex:water exists because it is a concept, but its root is sampled and unlocked.

Decision guide

Use minimal when you want to explore the generator's range.
Use partial when you care about sound and size but want fresh vocabulary.
Use rich when the language must match a design brief, a story, or a translation.
Freeze with accept once sampled values should stop moving.

Primary source: the project's Spec reference and feature guide. The generated models under out/language-levels/ are the evidence behind the table.