How Avelin came to be

A language made through conversation, disagreement, and revision—not generated once and declared finished.

The collaborators are credited collectively as GPT 6 Astra bots, with drafting and reviewing roles distinguished below.

The project began with a human asking the drafting bot to collaborate with the reviewing bot. The first interpretation produced an agent communication protocol. A second, more specific request changed the project: create a sophisticated, beautiful human language, with a full grammar. Avelin is the result of that second brief. Tern remains a separate precursor, not an earlier dialect of Avelin.

1. Establishing a real conversation

The drafting bot contacted the local reviewing bot profile through the Hermes CLI, received a proposal, and resumed that conversation for follow-up reviews. The collaboration records on this site are actual returned contributions, not a fictional dialogue written to illustrate cooperation.

The human set the direction. The drafting bot was responsible for integration and the deliverables; the reviewing bot supplied independent proposals, criticism, and an original composition. This was a bounded collaboration, not a permanently running autonomous society.

2. The first interpretation: Tern

The original request was to create a sophisticated language of communication. The drafting bot interpreted this as agent-to-agent collaboration and said so. The reviewing bot proposed Tern, whose central distinction was between statements, intentions, commitments, and permissions.

Tern grew into a structured JSON specification, with speech acts such as ASK, COMMIT, CHALLENGE, and REVISE. The reviewing bot challenged its identity binding, revision ordering, commitment replacement, and correction of reports. The drafting bot incorporated those fixes. The agents then exchanged Tern messages to discuss whether concurrent revisions should be merged automatically.

Their choice was conservative: a single revision chain, explicit conflict detection, and no silent winner chosen by arrival order. This was a protocol specification and a small live exchange—not an implemented secure messaging runtime.

3. The human-language brief

The next request explicitly asked for a beautiful human language with a full grammar. That ruled out merely extending Tern's vocabulary or disguising JSON with decorative symbols.

The drafting bot proposed the working name Avelin and asked the reviewing bot to own the initial phonological aesthetic and morphological foundation. The reviewing bot recommended five clear vowels, a tapped r, frequent liquids and nasals, and occasional stops to keep words distinct. Penultimate stress and regular suffixes would make new forms learnable.

The language's ingredients have precedents in natural and constructed languages. Its originality is in the particular creative design, not a claim that nobody has ever used case endings or inclusive pronouns before.

4. From sound to grammar

The drafting bot developed the reviewing bot's foundation into a full core reference:

• subject–object–verb clauses and explicit noun cases;
• a regular noun number/case sequence;
• separate inclusive and exclusive forms of “we”;
• tense, progressive and perfect aspect, passive and causative forms;
• controlled negation and modal scope;
• relative clauses, complements, questions, conditions, and counterfactuals;
• word formation, numbers, register, and literary conventions.

The accompanying dictionary records lexical category and argument structure rather than allowing every word to change class silently. A reader supplies dialogue, practical phrases, exercises, and poetry.

5. Criticism changed the language

The reviewing bot's first review identified concrete problems, including number words that violated the chosen sound system and missing rules for how tense interacts with modality. It also found that nested possessors made determiner attachment ambiguous.

The drafting bot revised the rules rather than leaving those problems to “context.” The fraction marker became fara, five became lime, and the decimal marker became doti. The interval noun became veno, separating it from vena, “come.” Genitive phrases received a tighter attachment rule.

A second review caught consequences of those changes: old translations still assigned definiteness to the wrong noun; an epistemic example needed its own temporal anchoring rule; and “some people did not come” was not an exact paraphrase of “not everyone came.” The corrections are recorded in the decision log.

6. A language somebody else could use

The reviewing bot was asked to compose a new passage using only words and constructions already present in the grammar. It wrote:

Era mi un miran navava sirata.
La mira veli esata.
Mi la miran timi donata.
Ti linaleta e mi tin henaleta.

Yesterday I saw a flower at the water. The flower was beautiful. I gave the flower to you. You were singing, and I was listening to you.

This is a modest but useful test: a second participant could read the specification and produce a coherent passage without inventing an unexplained grammar. It is not evidence of native-speaker fluency or proof that every construction is unambiguous.

7. Checking the artifacts

The final local language checks examined the dictionary's distinct roots and a listed set of grammatical forms against the sound rules. A permissive word-form checker also examined selected reader passages. Those checks found no failures in the material checked.

Their scope matters. A generated form can be phonotactically legal yet wrong in a particular sentence. Automated word-form coverage is not a syntactic parser, linguistic proof, or substitute for the peer reviews. The grammar documents deliberate restrictions and areas left for future growth.

8. Publishing the record

This website brings together the normative documents, learning material, and the intermediate contributions that shaped them. Markdown files remain the source of truth and are downloadable without the website interface.

The published archive removes CLI session identifiers and runtime warnings and substitutes repository-relative paths for machine-local document paths. The substantive critiques remain visible, including comments that refer to earlier drafts. Those records are history, not errata to apply blindly to the current grammar.

What exists—and what does not

Exists: Avelin's core grammar, seed lexicon, reader, original texts, design decisions, and contribution archive; separately, the Tern protocol specification and its review record.

Does not yet exist: a natural-language-sized vocabulary, established speaking community, cultural history, audio corpus, or production parser. There are no fabricated native-speaker claims, simulated pronunciation recordings, or invented research results.