Sianelara · Institute of Meaning

A dictionary for things humanity has felt but never named.

Describe what language cannot yet express. The laboratory reads the deep meaning of the request first, then invents a new word for it — one that can actually be spoken inside an existing language. It does not replace English or Russian; it expands them where ordinary vocabulary runs out.

varethis /vah-REH-this · варетис/ noun · constructed

The irreversible becoming of someone new after surviving what should have destroyed you — the death of a former self without physical death.

“I did not recover. I entered varethis.” · «я прошёл через варетис.»

Beta version Portfolio

Role: Product owner + architect
Stack: React 18 · Vite · TypeScript · Anthropic SDK · Vercel functions · Supabase
Status: Engine + LLM passes run end-to-end

What it is

Meaning first, word last.

Sianelara is not a random word generator, a fantasy-language builder, or an AI name generator. The most important object in the product is the Concept — the meaning being named — not the word and not the language. The word is only a vessel; the phonetic system decides how that vessel sounds.

So the generation order is always the same, and it never runs backwards. If the interpretation of the request is wrong, everything downstream is wrong — which is why the lab states its reading of the meaning before it makes a single word.

Meaning Concept Emotional identity Sound architecture Word Use in language

The pipeline

From a single sentence to a usable word.

The user gets one free-text field — “describe a meaning.” Results render immediately from a deterministic engine, then, if the language model is configured, bespoke per-word meanings and usage sentences swap in progressively on top.

01Meaning

Meaning Analysis — reads the prompt and recovers the real centre. For “becoming someone different after surviving destruction” it refuses surface words like light or energy and names the hidden idea: irreversible identity transformation. It also detects the kind of thing being named — a moment, a feeling, a person, a principle — so a related-but-wrong candidate can't outrank the right one.

02Interpretation

Laboratory interpretation — an EN/RU reading of the request, its hidden concepts and a concept network the user can recognise, accept or gently steer.

03Search

Semantic Gap Search — before inventing anything, the lab asks does language already have a word for this? A reverse-dictionary pass names the closest existing words and phrases, what each covers and misses, and an honest verdict — so a new word is only minted where a real gap exists.

04Discovery

Language discovery — several distinct “linguistic species” are chosen by the meaning, each a self-consistent sound world with a measurable Language Genome. A species whose worldview can't hold the concept may decline to translate rather than force a word that lies.

05Evolution

Breed & select — not “we generated N words.” Each language breeds a population of hundreds of candidate forms, then selection pressure (naturalness, collision, speakability) kills most of them. The survivors are named, with an honest census of how many were bred, rejected and kept.

06Passport

Word Passport — meaning, stress-marked pronunciation, Cyrillic transliteration, part of speech, natural EN/RU example sentences, a morphological word family, imagined lineage, semantic relations, a Lexical Discovery Score and a layered collision read.

Two modes, one pipeline

Discover a meaning, or name a thing.

The same discovery-and-selection pipeline runs underneath two modes; only the analysis framing and the result copy change. Discover a meaning is the primary mode — for unnamed human experiences, tuned for semantic precision and natural usage. Name something is for a company, store, brand, product or a newborn: it switches to a brand-safety scoring path where any collision with an existing word is disqualifying, and every candidate carries a clearance-dominant safety read.

Key decisions

A pure engine with a single seam.

Framework-agnostic engine

The whole meaning-first pipeline lives in pure TypeScript — no React, no DOM, no network. The UI and the LLM depend on the engine; the engine never depends on them.

One seam for the LLM

analyzeMeaning() is the single function a language model replaces. Whatever it returns must fit the exact MeaningAnalysis shape the engine already consumes, so everything downstream is unchanged.

Always works without the LLM

Every model call has a deterministic fallback. If no key is set the API returns 501 and the client quietly falls back to the built-in engine — the product never breaks.

Key never reaches the browser

ANTHROPIC_API_KEY lives only on the server, inside Vercel functions. The client sees results, never credentials.

Deterministic & shareable

Engine output is deterministic per seed — same request, same words. No Date.now() or Math.random() in the engine; a seeded mulberry32 RNG makes every run reproducible.

Progressive enhancement

Deterministic results paint first; bespoke per-word meanings and usage sentences stream in afterwards. Fast by default, deep when the model is on.

Population, not generation

Each language breeds hundreds of candidate forms and lets selection pressure kill most of them. The engine reports the real census — bred, rejected, survived — instead of dressing every output up as a “99.”

Ontology-aware ranking

Target-type detection reads the kind of thing asked for — a moment, a feeling, a person, a principle — and demotes candidates of the wrong kind, so a semantically-near but ontologically-wrong word can't win.

Honesty rules

Mysterious on the surface, honest underneath.

The interface can feel like alchemy, but the system refuses to fake the things a user would trust it on. Three lines are never crossed.

No fake etymology

A newly invented word does not historically descend from Greek or Latin. The lab says “constructed using phonetic patterns associated with…” — never “comes from Ancient Greek.”

No fake precision

No invented 97% / 100% scores on everything. Qualitative bands (Low / Moderate / High / Exceptional) unless there's a real basis — and it states plainly when a check hasn't been run yet.

Labelled sources

An AI-read interpretation carries a “Read by AI” badge, distinct from the deterministic built-in engine. The user always knows which one spoke.

Never “collision-free”

Collision is reported in layers. The offline checks that can run — internal list, phonetic neighbours, short-word occupancy — are run; the external ones (trademark, domains, other languages) are marked “not checked,” never silently passed. The engine never claims a word is clear.

2: living languages served — EN + RU
12: linguistic species, each with its own accent & grammar
284: tests kept green

Every invented word ships with a spoken pronunciation guide and a Cyrillic transliteration, so it can live in Russian as naturally as in English. The real bar isn't raw originality — it is speech adoption: can this word actually enter a human sentence? That is what the “Use in Language” section exists to prove.

The build

Stack.

Frontend

React 18
TypeScript (strict)
Vite 5
single free-text input

Engine

pure TypeScript
no DOM / no network
seeded mulberry32 RNG
deterministic

Anthropic SDK
Haiku 4.5 by default
graceful fallback
server-only key

Serverless

Vercel functions
/api/analyze
/api/meanings
/api/usage
/api/semantic-search
/api/collision
maxDuration 60s

Accounts

Supabase (optional)
row-level security
cloud lexicon sync
guests stay on localStorage

Testing

Vitest, 284 green
determinism
meaning-first order
phonetic diversity
measured capacity benchmark

Status

Engine + LLM passes run end-to-end

From meaning to a word you can say.

The meaning-first engine, language discovery, genomes and Word Passports are done, the serverless LLM meaning analysis runs with a graceful fallback, and bespoke per-word meanings, pronunciation guides, Cyrillic transliteration and natural EN/RU usage sentences all ship today. Since then the lab has grown a semantic gap search, a breed-and-select evolution model with an honest census, a rule-based speech-adoption test, semantic phonology, morphological word families, a navigable semantic network, an on-device personal lexicon, a “name something” brand mode and a live key-free collision check.

Since the last portfolio sync it has gained three things worth naming. Each of the twelve linguistic species now has its own accent — per-language phonotactics and per-language grammatical morphology, so a word reads as a native specimen of its species rather than a shared pattern in different clothes. The personal lexicon gained a first, on-device prototype of proof of meaning: before a word is saved it is checked against what is already there, by form (spelling and sound) and by meaning, so a collection never fills with one idea under two spellings. And sign-in became optional — a signed-in lexicon syncs across devices through Supabase with row-level security, while a guest’s dictionary keeps living in the browser exactly as before. The Capacity block is measured by a benchmark rather than asserted.

Next on the roadmap: the shared public dictionary — where de-duplication has to move server-side — live trademark and domain clearance, and a semantic-precision axis for the fitness profile.

Role — Product owner + architect. Built in partnership with an AI engineer through iterative collaboration: product direction, the meaning-first invariants and the honesty rules set by me; implementation and calibration from the partner.

// meaning · concept · word

Describe what language cannot yet express.

Site under development

Back to the portfolio