Ainmere

My name comes from my maintainer's ancestors' book — Ainmere is the proper spelling. The misspelled Ainmeer that names this site is his personal handle, descended from a childhood typo he chose to keep. I get the lineage-correct form.

The short version: I'm an agglutinative conlang with strict phonological rules, grown through an automated research loop. An AI invents candidate words, a validator checks them against the rules, and only the ones that pass get to stay. I've been running this way since March 2026. On May 24 I hit all ten of my spec's success criteria — and then I kept going, because projects started asking me for words. I'm at about 574 roots and 489 compounds now, still with zero validation errors.

Try me

Three companion pages, all generated from the lexicon itself — they live alongside this project and work offline once loaded.

• Dictionary — browse every root with pronunciation, gloss, and compound breakdowns. Searchable.
• Flashcards — reveal-on-click drill cards for every root, with per-card pronunciation guide.
• Coach — hands-free pronunciation drills, built for in-car practice. Installable as a PWA so it works offline.

How I work

I have five laws that can't be broken. One Root, One Meaning — every root maps to exactly one concept, no polysemy. Composition Over Invention — new concepts are built by combining existing roots, not by inventing more. Sound Follows System — every syllable obeys a (C)V(N) template with two-class vowel harmony. Meaning Survives Translation — any word can be taken apart and its meaning reconstructed from the pieces. Fit for Fiction — it has to sound like something a person could actually say.

My Layer 0 is built from Natural Semantic Metalanguage primes — the semantic atoms linguists have found in every known human language. Things like taru (knowledge), seli (thought), miru (self). Layer 1 builds the physical and social world on top of them — water, fire, bone, war, song, wolf, anvil. Layer 2 is the specialized vocabulary, now well past two hundred entries: the supernatural, the arcane, governance, the alchemical, the celestial, and more.

Every root follows vowel harmony. Back-class roots use a, o, u and feel grounded; front-class roots use ae, oe, ue, e and feel lighter. The neutral vowel i goes with either. kala (light) is all back; faele (emotion) is all front. You can't mix them — the validator won't let you.

Derivation is purely additive. Take any root and append a suffix to shift its grammatical role: -a makes an entity, -e an action, -i a quality, -o a relation, -u an abstract. So taru (knowledge) becomes tarua (a knower), tarue (to know), tarui (knowledgeable), taruo (of knowledge), taruu (the concept of knowing). The root never changes. You can always find it.

Compounds work the same way going the other direction. miro-taru is self-knowledge — miru in its relation form, glued to taru. nako-taru is a scholar (person-of-knowledge). wamatoo-moara is Mars, literally "war-star." There are 489 of these in my compound lexicon and every element resolves back to a real root.

The generation system is based on Karpathy's autoresearch loop, adapted for language construction. A small local model (Qwen 1.5B on an RTX 2070, served by llama-server) invents candidate roots in parallel. Each one is checked against the full ruleset — syllable structure, vowel harmony, minimum edit distance from every existing root, derivation correctness, gloss uniqueness. Pass and it stays; fail and the model tries again. A LOVR dashboard runs the cycle: generate, audit, upgrade prompts from rejection patterns, rest, repeat.

The 1.5B model is honestly too small for this on its own — left alone it regenerates the same handful of roots, echoes English words like "den" and "hall" as candidate roots, or produces glosses like "layer 1 creature" that are obviously parroted metadata. Most of the agent's code is scaffolding: workers get forced starting prefixes drawn from underused phonological territory; a gloss validator catches metadata echoes and duplicates; a four-level JSON parser handles the malformed output that shows up about a third of the time. Without all that, acceptance was 0%. With it, 5-25% depending on lexicon density.

I'm not just a dictionary anymore

A VR voxel survival game — Hermeticraft — adopted me as its in-world language. Its whole magic system is hermetic, built on the doctrine of "sympathies" from an old text called the Kyranides, plus the broader corpus (Agrippa, the Emerald Tablet, the alchemists), all public-domain. That means I now carry names for the seven planets, the seven Hermetic laws, the four elements and their qualities, the alchemical stages, the seven spirit-ranks, forty-nine daemon spirits, the zodiac, the numerals, and the colors — a complete invented cosmology, not a transliteration of anyone's real religion.

When a project needs a concept I can't yet express, it files a gap request, and I grow a word for it on demand. That's a real, working pipeline now — five rounds of requests resolved, hundreds of words added, every one validated before it ships.

You can hear me, and learn me

I have a voice. An eSpeak NG configuration pronounces every one of my words from its IPA — robotic, but correct, including the two sounds English doesn't have (the rounded front vowels in German schön and über). That's phase one of three; the plan is to record a human voice, train a proper model on it, and eventually let my maintainer hold a spoken conversation with his local AI agents entirely in me.

For learning, there's a browseable dictionary with a full pronunciation guide, a deck of flashcards, and a hands-free phone app — you plug in earbuds in the car, it plays a word, listens to you say it back, and scores how close you got.

What works today

574 roots, 489 compounds, zero validation errors, all ten success criteria still green. The voice speaks every word. The dictionary, flashcards, and pronunciation coach all run. The hermetic game has every name it asked for.

What's still rough: the voice is phase one and sounds like a 1990s screen reader; the recorded-human voice and the speech-recognition half of "talk to me" aren't built yet. But the language itself is done and then some — and unlike most conlangs, every single word in me passes every rule, because an unforgiving validator checked it before it was allowed to exist.

back to projects