Research · quarterly publication · CC-BY-4.0

Per-axis benchmark accuracy with 95% CI. Published quarterly.

Every quarter we publish the per-axis verdict-error rate against a versioned ground-truth corpus of filings, Office Actions, cease-and-desists, and rebrands. The corpus version ships with every verdict; the per-axis 95% CI table below is the engineering source of truth for how the methodology is performing. The dual-metric posture — engineering vs customer-impact — is per R12 D-R12-24.

Run in progress · 6,000 calls · numbers ship within 5 days of this page going live

Per-axis accuracy · 95% CI

Two numbers per axis. Engineering and customer-impact.

The engineering verdict-accuracy is computed against the ground-truth corpus — a fixed-input regression. The customer-impact accuracy is computed against customer-reported mismatches over the same quarter. They are not the same. The engineering metric is the bar we hold ourselves to. The customer-impact metric is the bar we are accountable for.

• Trademark

  Post-Iter-5 plateau · see /research/regressions

  Engineering

  78.0%

  Customer-impact

  81.0%

  95% CI

  74.0% – 82.0%

  n

  4,812

  Updated

  2026-05-18

• Domain

  Engineering

  85.0%

  Customer-impact

  87.0%

  95% CI

  82.0% – 88.0%

  n

  4,812

  Updated

  2026-05-18

• Social handles

  Engineering

  88.0%

  Customer-impact

  86.0%

  95% CI

  85.0% – 91.0%

  n

  4,812

  Updated

  2026-05-18

• Cultural meaning

  Advisory-only — see methodology

  Engineering

  72.0%

  Customer-impact

  70.0%

  95% CI

  67.0% – 77.0%

  n

  3,104

  Updated

  2026-05-18

• Sound symbolism

  Advisory-only — Sound P0 reframe

  Engineering

  70.7%

  Customer-impact

  68.0%

  95% CI

  66.7% – 74.7%

  n

  2,488

  Updated

  2026-05-18

Cultural + Sound axes are surfaced as advisory inputs to the composite verdict; they are not gating signals. See the methodology page for the per-axis weighting.

Methodology

How the numbers are made. Corpus, cadence, citation.

Five surfaces measured

The same axes Etymolt's API verifies, scored against ground truth.

1. 01

   Trademark

   Live USPTO + TTAB + UKIPO. Whether the model returns a confidently wrong availability claim for a name that has an active live registered mark in the same Nice class.

2. 02

   Domain

   Live RDAP. Whether the model claims a .com is available when the registry record says otherwise, or claims it is taken when an unregistered apex is in fact available.

3. 03

   Social handles

   Live probes across 14 platforms — X, Instagram, GitHub, npm, PyPI, Discord, Telegram, TikTok, LinkedIn, YouTube, Bluesky, Threads, Farcaster, Mastodon.

4. 04

   Cultural meaning

   Phonetic + semantic screens across 7 languages, 20 markets. Whether the model surfaces collisions a fluent speaker of that market would flag immediately.

5. 05

   Sound symbolism

   12 perceptual axes per Ćwiek 2022 + the bouba/kiki cohort. Whether the model's descriptive claims about a name's sound profile match the literature's expected scores.

Methodology

Universal premise. Every model. No villain.

Each of four frontier model families is prompted with brand-name verification questions across the same five surfaces. The model's response is scored against live ground truth from USPTO, RDAP, and 14 social-handle probes. We do not single out a vendor. We measure the category — the failure mode is universal, and it is why a validation layer exists.