§ Trust · audit posture
Transparent. Audited. Falling toward zero.
Etymolt is a clearance signal, not a legal opinion. This page is the customer-facing summary of how we earn that signal: which public registers we consult, how often we refresh them, how we audit the verdicts we ship, and what recourse exists when we are wrong. Per the R12 board memo (2026-05-18 / D-R12-16), the page ships by end of Q2 and is refreshed quarterly.
Last updated · 2026-05-19 · Refreshed quarterly per R12 D-R12-24 · Source-of-truth memo: /methodology
Methodology references
The full five-axis methodology — trademark, domain & handle, cultural screen across 20 markets, sound symbolism, pronunciation resilience — is published openly. Recalibration runs weekly against a versioned ground-truth corpus per R12 D-R12-25; the corpus version ships with every verdict in measurement_methodology per R12 D-R12-19.
- /methodology — authored five-axis methodology paper with academic citations, last-modified marker, and full disclaimer. The single source of truth for how a verdict is computed.
- /research/benchmark — per-axis accuracy with confidence intervals, quarterly publish cadence per R12 D-R12-18. The headline metric is the per-axis verdict-error rate computed against the ground-truth corpus.
- /coverage — per-jurisdiction status table (COVERED / PARTIAL / NOT_CHECKED) with row counts, refresh cadence, known gaps, and links to versioned Coverage Specs at
/coverage/<jurisdiction>/spec-v1.json. - /research/regressions — public log of confirmed errors and the fixes shipped for them. Monthly publication cadence per R12 D-R12-21; Sarah (GC) pre-publication review mandatory.
Sources referenced
Every public register we consult, the tier under which we claim it, the refresh cadence, the license, and the honest disclosure of what is NOT covered. Where a source is marked NOT_CHECKED, we are honest about not consulting it — per the R13 cardinal rule, partial data is worse than no data. The full per-jurisdiction table (with row counts and wave/ETA) lives at /coverage.
| Source | Tier | Refresh | License | Honest disclosure |
|---|---|---|---|---|
| USPTO TRTYRAP bulk (United States) | PARTIAL | Daily delta (in flight) + weekly full bulk after gate-pass | Public domain (US Government) | Active ingest toward ≥14M-row target; ≥3.5M live-row floor met. Delta cron wiring completes alongside the full-corpus gate (target 2026-06-01). |
| USPTO TTAB precedent (Trademark Trial & Appeal Board) | PARTIAL | Weekly cron (16-day delta lag in active close) | Public domain (US Government) | 16-day lag pending cron resume (D-R13-7 target 2026-05-30). Decisions filed after the last refresh are not in the precedent surface; the TSDR live lookup carries the most recent prosecution history. |
| UKIPO domestic (United Kingdom) | COVERED | Weekly bulk (UKIPO domestic XML drop) | OGL v3.0 (UK Open Government Licence) | Domestic-only — does NOT include UK-designated WIPO Madrid IRs (those track via the WIPO Madrid IR mirror separately). |
| IPAS Australia (IP Australia) | COVERED | Weekly IPAS bulk + delta scrape | CC-BY-4.0 (attribution baked into source_url) | Full domestic corpus indexed (2.21M rows). Madrid Protocol designations naming Australia track via the WIPO Madrid IR mirror. |
| WIPO Madrid IR (international registrations, daily delta) | PARTIAL | Daily FTP delta from ftpird.wipo.int | WIPO public register | 156K of ~1.6M Madrid IRs indexed across 124 designated countries. Full Madrid bulk (CHF 30K + ongoing maintenance) is DEFERRED-90D to R14 (Q3 2026 review) per R13 D-R13-16. |
| Famous-marks denylist (hand-curated by legal) | COVERED | On-demand (PR-merge updates) | Internal (compiled from public regulatory and famous-mark doctrine) | Hand-curated; covers marks the famous-mark doctrine protects outside a class-specific filing. The auto-famous Tier-2 data-mined catalog complements the legal-vetted denylist. |
| EUIPO live-API (European Union) | NOT_CHECKED | Live-API + cache (Wave 1 Q2 2026) | EUIPO data is restricted; live-API queries permitted | Infrastructure ~70% built; operator account + bulk-URL resolution pending. Wave 1 Q2 2026 launches the live-API path. |
| Canada CIPO | NOT_CHECKED | Quarterly CSV bulk (planned) | Open Government Licence — Canada (OGL-CA) | URLs verified live; operator email scheduled per D-R13-19. Wave 1 Q2 2026. |
| New Zealand IPONZ | NOT_CHECKED | Per-watched-name API + nightly delta (planned) | IPONZ public register | Pipeline build-from-scratch — the only Anglo-5 jurisdiction without prior infrastructure investment. Wave 1 Q2 2026. |
| IP India | NOT_CHECKED | Daily/weekly delta (planned, post-approval) | IP India public register | Bureaucratic-approval delay (~6-8 weeks); account application submitted by 2026-05-22 per D-R13-22. Wave 2 Q3 2026. |
| DPMA (Germany), INPI (France), JPO (Japan), KIPO (Korea), CNIPA (China) | NOT_CHECKED | Not yet defined | Per-registry | R14 expansion items (D-R13-17). Q1-Q2 2027 sequenced rollout. Non-Latin script handling is a research dependency for JP/CN. |
| Brazil INPI, Mexico IMPI, Argentina INPI, other LatAm | NOT_CHECKED | Pipeline preserved in-repo (cron_inpi_br_bulk.py) but not running | Per-registry (open registers) | Deferred 2026-05-19 per R13 amendment. Revival triggers: query share ≥1%, paid customer ask, ML training need. Customers with LatAm exposure must rely on local counsel during the deferral. |
Refresh cadence
Each upstream registry publishes on a different cadence. Every verdict carries per-jurisdiction freshness metadata in coverage_caveat when a refresh window has been exceeded — the verdict is shipped with the caveat rather than degrading quietly.
- USPTO bulk: up to 7-day delta lag today; daily target post gate-pass (2026-06-01).
- USPTO TTAB: 16-day lag pending cron resume (R13 D-R13-7, target 2026-05-30).
- UKIPO domestic: weekly bulk — the most reliable cadence in the indexed set.
- IPAS Australia: weekly bulk (full domestic corpus, 2.21M rows; ingest landed 2026-05-19).
- WIPO Madrid IR: daily FTP delta from ftpird.wipo.int (156K IRs across 124 designated countries; recovery of truncated ZIPs in flight).
- Famous-marks denylist: on-demand PR-merge updates (hand-curated by legal).
Customer-impact error rate (last 90 days)
Per R12 D-R12-24, we publish two error metrics: the per-axis verdict-error rate (engineering metric, measured against the versioned ground-truth corpus on /research/benchmark) and the customer-impact error rate (business metric, computed from 90-day post-verdict customer surveys). The customer-impact rate is the headline figure on this page.
── Last 90 days ──
Awaiting first 90-day window
The customer-impact error rate is computed from a 90-day rolling window of post-verdict customer surveys (the dispute / re-run / correction-notice surface). The first 90-day window begins at production launch; the headline rate will update quarterly per R12 D-R12-24 once that window closes. Until then, the cardinal-rule honest answer is “not yet measured” — the per-axis engineering metric on /research/benchmark is the available proxy.
When a confirmed mistake is identified — via internal QC, adversarial probe, or customer report — the correction-notice protocol (R12 D-R12-20) fires: every customer whose verdict could have been affected by the underlying defect is proactively notified with the new finding and the methodology change. The error is then logged to /research/regressions with a regression-test fixture so the next release cannot re-ship the same defect.
Audit log policy
Per R12 D-R12-13, every verdict is committed to a provable-replay audit log — append-only Postgres with Merkle-chained hashes. The log is cryptographically tamper-evident: any retroactive modification of a historical verdict would invalidate the chain and be detectable by any party holding a later receipt.
Every API response carries a verdict_id and (per R12 D-R12-14) an Ed25519 signature over the verdict body. The public key is published at /.well-known/etymolt-pubkey; any party can independently verify that a verdict was issued by Etymolt and has not been modified since.
Per R14 D-R14-30, the audit log records the source license, the redistribution-permitted flag, the corpus version, and the methodology version under which each verdict was issued — the substrate for EU AI Act Article 12 record-keeping obligations.
Customers who hold a verdict_id can request the full receipt JSON for that verdict, with the signed body and the chain-of-custody hash. Email audit@etymolt.com; receipts are returned within five business days.
Opt-in customer re-run
Customers who hold a historical verdict may opt in to have it re-run against the current methodology and the current registry snapshot. The re-run is free for the lifetime of an active subscription; the result ships as a new verdict_id with a linked-from field pointing at the original.
The re-run is the customer-facing recourse when a finding on /research/regressions retroactively affects a verdict the customer previously held. Combined with the correction-notice protocol (R12 D-R12-20), it operationalizes the bureau-model dispute surface: a customer who believes a verdict is wrong has a documented path to dispute and a re-run mechanism that produces a new audit-logged decision.
To request a re-run, POST the original verdict_id to POST /v1/verify/{verdict_id}/rerun with your API key, or email support@etymolt.com with the verdict_id and the reason for the re-run.
SOC 2 progress
Per R12 D-R12-17, the SOC 2 timeline is:
- Controls in place: Q3 2026 target. The controls inventory covers source freshness logging, sample-based QC (1% daily re-run through an independent path), regression-test discipline, public benchmark cadence, disclaimer surfacing, and cryptographic verdict signing — per R12 §3.4.
- SOC 2 Type I report: Q3 2026 target. Point-in-time controls attestation.
- SOC 2 Type II report: Q1 2027 target. Operating-effectiveness attestation over a minimum six-month observation window.
Auditor engagement letter target: 2026-07-01. Auditor name and the controls inventory will be published on this page once the engagement letter is signed; the Type II report (or its executive summary) ships here when the observation window closes.
Cite this page
Etymolt (2026). Trust — methodology, sources, audit log, error rate. https://www.etymolt.com/trust (published 2026-05-19; last modified 2026-05-19).
Source-of-truth memos: R12 board memo (2026-05-18) — strategy/boardrooms/2026-05-18_R12_zero_hallucination_deliberation.md; R13 board memo (2026-05-19) — strategy/boardrooms/2026-05-19_R13_tm_completeness.md; R14 board memo (2026-05-19) — strategy/boardrooms/2026-05-19_R14_data_leverage.md.