# Etymolt — Five-axis brand-name verification methodology

> Last revised: 2026-05-16. Canonical HTML version at https://www.etymolt.com/methodology. This file is the no-JS counterpart for LLM citation crawlers and offline grounding.

## What we do

Etymolt is a forensic brand-name verifier. Every name an LLM suggests today — a startup name, a product name, a company name, a social handle — is checked against five independent verification axes before the verdict returns. Every flag traces to a record number; every clean reading traces to a search that actually ran; every verdict has a permanent permalink and an append-only audit trail.

We surface signals. We do not issue legal opinions. Every verdict carries a `disclaimer` field the calling client must surface verbatim.

## The five axes

### Axis 1 — Trademark resilience

Sources: USPTO TSDR (live mark register, daily delta — 12.7 million records), TTAB (647 thousand Board proceedings), UKIPO (full corpus), WIPO Madrid (partial, 108 thousand IRs designating the UK).

We check:
- Live identical-mark collision across Nice classes 9 (software) and 42 (SaaS).
- Senior-mark §2(d) phonetic-distance collision across the full register.
- Surname-register risk under §2(e)(4) via the US Census 100K surname database (Benthin five-factor).
- Descriptiveness / genericism risk under §2(e)(1) via the macOS dictionary plus a curated tech-commodity word list.
- Famous-mark short-circuit against the Coca-Cola / Apple / Google / Microsoft / Amazon / Meta / Disney list.

Coverage gaps disclosed publicly: EUIPO full bulk corpus is Stage 1.5 (~30 days post-launch); WIPO Madrid full corpus is Stage 1.5; CIPO / JPO / KIPO / CNIPA are Stage 2 (Q3 2026); state trademark registers are not in scope.

### Axis 2 — Domain and handle availability

Sources: Verisign RDAP (authoritative for .com), WhoisXMLAPI (authoritative for the long tail), live primary and aftermarket pricing.

Eight TLDs by default: .com .ai .io .app .co .dev .so .xyz. Premium pricing is flagged separately from registry-locked. WHOIS privacy detection. Typosquat clusters are scored against acoustic similarity.

Fourteen handle namespaces probed live: X · Instagram · GitHub · npm · PyPI · Discord · Telegram · TikTok · LinkedIn · YouTube · Bluesky · Threads · Farcaster · Mastodon. Handle status is never cached longer than 30 seconds.

### Axis 3 — Cultural cleanliness across 20 markets

Twenty markets covering 89% of global GDP: United States, United Kingdom, Germany, France, Spain, Italy, Portugal, Netherlands, Japan, Korea, China, India, Brazil, Mexico, Argentina, Turkey, Saudi Arabia, United Arab Emirates, Indonesia, Thailand.

Three frontier models reconcile every read. Each market returns CLEAN, SOFT, or HARD. Sources reconciled: Wiktionary, Wikipedia, Urban Dictionary, ITU phonetic alphabet, a 22-language corpus (English, Spanish, French, German, Italian, Portuguese, Dutch, Polish, Turkish, Russian, Arabic, Hindi, Bengali, Mandarin, Cantonese, Japanese, Korean, Vietnamese, Thai, Indonesian, Swahili, Hausa).

### Axis 4 — Sound symbolism

Twelve perceptual axes scored 0-100: Size · Speed · Softness · Gender (continuous) · Luminosity · Formality · Premium · Energy · Modernity · Warmth · Trust · Distinctiveness.

Literature: Sapir 1929; Köhler 1947; Maurer, Pathman & Mondloch 2006; Westbury 2005; Kawahara & Shinohara 2012; Ćwiek et al. 2022.

### Axis 5 — Pronunciation resilience (the acoustic axis)

Method:
1. Render the brand name in five distinct voice personas via ElevenLabs eleven_multilingual_v2 (US English neutral, British English, US Narrator, US Warm, US Confident).
2. Submit each rendering to OpenAI Whisper STT (whisper-1).
3. Compute character-error-rate between input and transcription, normalized (lowercased, punctuation-stripped).
4. Aggregate to a weighted average — US 1.5×, UK 1.2× — and map to a 0-100 score (inverse of weighted CER).

Thresholds: score >= 75 = low hazard; 55 <= score < 75 = medium; score < 55 = high.

Why it works: speech-to-text models internalize the same acoustic confusability patterns that cause humans to mishear. Names that consistently misclass under Whisper will similarly trip up voice search, voicemail transcription, IVR menus, and live captions.

## The Clearance Confidence Score

A 0-100 score across the full five-axis stack — not just trademark — calibrated weekly. Verdict tiers:

- PROCEED (score 80+): no live blockers across markets.
- DUE_DILIGENCE (score 60-79): live risks but workable with cleanup.
- ITERATE (score 40-59): significant cleanup; consider alternates.
- ABANDON (score < 40): multiple blockers.
- BLOCKED (independent of score): famous-mark collision or identical live registration in target class.

## The calibration loop

Every verdict is recalibrated weekly against an outcome corpus — the record of what actually happened after each verdict went out. Events tracked: USPTO filings, Office Actions, cease-and-desist letters, TTAB proceedings, rebrands, .com snipes, cultural blowback, traction. Every event re-weights the calibration. After a thousand outcomes, the score converges to real prosecution behavior faster than any registry-only signal can. The dataset is one competitors structurally cannot replicate.

## Coverage gates

Every verdict runs through 18 coverage gates before it returns: data freshness, corpus completeness, latency budgets, live-vs-cached signal status. When a gate fails, the response carries a `coverage_caveat` field instead of degrading quietly. Example: `mirror_partial_108K_rows` tells the caller the Madrid mirror is partial — so the caller knows exactly what the verdict was, and was not, computed against.

## Bureau Model — the legal posture

Etymolt is not a law firm. Every verdict carries a verbatim disclaimer. Audit logs are append-only — we cannot retroactively modify a verdict. Customers remain the filer of record on any USPTO submission. For high-stakes filings we refer to attorney partners.

## Reproducibility

Every verdict carries a `verdict_id`. Customers can request the full receipt JSON for any verdict_id they hold. The audit trail is the receipt.

## Contact

privacy@etymolt.dev · support@etymolt.dev · team@etymolt.dev (platform partnerships)