The verdict is a received object.
A short book about how Etymolt builds its verdicts, what each axis means, where the model's confidence ends, and why the disclaimer is the most important field in the response.
What we mean by a verdict.
A verdict is not an opinion. It is a structured statement of what the consulted registries said, traced to record numbers, signed at issue time, and ended with a recommended next action. The verdict is a received object — built to be handed off to an agent, a counsel, or a founder, and to survive scrutiny in each setting.
Most products in this category answer the question "is this name available?" with a string. We answer it with an object. The object carries six things the string cannot: a status drawn from a closed enumeration of five values, a numeric score with a stated confidence interval, the volatility distance to the next status state, per-axis confidence with named uncertainty drivers, the coverage caveats per consulted source, and an Ed25519 signature over the payload digest.
The signature is not decorative. It exists so that a downstream agent can verify, client-side, that the verdict was issued by Etymolt and has not been altered between issuance and use. This matters when the verdict is going to be quoted in a Series-A diligence room six months from now, or when a customer's lawyer wants to confirm that the registry citation predates the trademark filing being argued about.
We do not claim the verdict is a substitute for trademark counsel. The Bureau Model — explained in Chapter 6 — sets the boundary: we report records of record, we do not opine on infringement. The verdict carries that boundary as a verbatim field on every response.
The verdict object — structure as receipt.
The schema is locked and versioned at EVP/1 — the Etymolt Verdict Protocol. It is published under CC-BY-4.0 so any third party can implement against it. The current version, v1.0, returns nine top-level fields. We walk through each in order of the question it answers.
verdict — one of PROCEED · ITERATE · DECIDE · ABANDON · INSUFFICIENT_SIGNAL. Each verb-phrase carries the implied next action; the status taxonomy is closed so that downstream code can pattern-match without surprise.
score — an integer between 0 and 100. It is the compressed read of the five axes combined under the calibration logic discussed in Chapter 4. The score alone is not actionable; the score with its confidence interval and volatility distance is.
axes — a map of five named fields, each carrying its own integer score and structured detail. The full axis taxonomy is the subject of Chapter 3.
axes_confidence — per axis, a confidence score and a named list of uncertainty drivers. "single-LLM derivation, no cross-model corroboration" is a string we report when we mean it. The point is that the verdict tells you why it doesn't know what it doesn't know.
confidence_interval — lower and upper bounds around the headline score, computed via axis_uncertainty_propagation_v1. The full method is published in the regression log.
verdict_volatility — the score at which the verdict would flip, and what it would flip to. If a verdict is one point from changing, you find out before you ship.
coverage_caveat — per jurisdiction, the freshness window and known unavailability of the underlying source. EUIPO mirror_not_loaded is a string we say when we mean it.
signature + signature_key_id — Ed25519 over the payload digest. Key rotation is documented under the EVP/1 spec at https://github.com/etymolt/evp-spec.
disclaimer — verbatim. Always present. Never elided by intermediate systems. The disclaimer is what makes the Bureau Model promise durable.
The five axes — each traced to a record.
The verdict's score is a compression of five axis-scores. Each axis is a separate question, answered by a separate data source, with its own method, freshness window, and confidence model. The full reference is at /coverage.
Trademark — USPTO TRTDXFAP and TTAB · EUIPO live API · UKIPO domestic · WIPO Madrid IR · famous-marks denylist. POCA orthographic and phonemic confusability. §2(d) likelihood-of-confusion modeled per Nice class.
Domain — six TLDs queried:.com.ai.dev.io via RDAP authoritative;.so.co via DNS fallback. Six handle platforms — GitHub, X, npm, PyPI, YouTube, TikTok — checked in parallel. Workaround variants surfaced as alternatives.
Cultural — two-tier. Tier 1 synchronous: 30-language Hurtlex denylist, Unicode-confusable fold, 67-entry disasters list, ~50 sacred-names catalog. Tier 2 asynchronous: a three-LLM advisory panel — Claude Haiku, GPT-4o-mini, Gemini — cached out-of-band and reconciled into the next-day score.
Sound — a static rule-based heuristic. Pure-Python phoneme-class rules and phonetic-neighbor distance. English ASCII orthography only on /v1/verify. The method is public. The axis carries advisory: true because it is not a learned model and we do not claim it is.
Pronunciation — on /v1/verify, a consonant-cluster heuristic, en-US only, advisory. On /v3/voice/hazard, the full Whisper round-trip pronunciation hazard scoring across six accents for production pipelines. The same axis, two different depths, chosen by the caller.
Calibration — what the verdict doesn't know.
The single feature that makes Etymolt's verdict structurally different from a chatbot's answer is calibration. Every score carries a confidence around it. Every axis tells you what would change its mind. Every verdict ends in a recommended action — not a vibe.
The methods we use for confidence-interval propagation, volatility computation, and uncertainty-driver attribution are published in the regression log at /research/regressions. The point of publication is not generosity. It is that a verification claim that can't be reproduced is not a verification claim.
Coverage caveats — the registry is not the world.
Every consulted source has a stated freshness window. EUIPO returns slower than USPTO. Handle platforms cache. The 30-language denylist has uneven density — fifteen of twenty-one listed markets have zero curated patterns. We say this because the verdict's credibility depends on its limits being named.
The full coverage spec is at /coverage. The point of the page is operational, not marketing — a developer integrating Etymolt should know which axes will fire on which markets before they ship.
The Bureau Model — why we don't opine.
Etymolt operates under the Bureau Model legal posture: we report what the records of record show; we do not opine on infringement. This is not a disclaimer pasted onto the bottom of a SaaS landing page. It is the architectural choice that makes the verdict object durable.
The verdict tells you what the records show. The decision to file, to adopt, or to abandon is yours and your counsel's. We make that decision easier by carrying the evidence — but we do not make it.