§ Changelog · v1.0 → present

Every release, signed and dated.

Bylined by the engineer who shipped it. Breaking changes flagged at least two weeks ahead. Deprecations announced before they happen. RSS available — subscribe and you'll know before your customers do.

Live · 7 releases since 2026-04 · RSS feed · EVP/1 versioning policy ↗
v2.2.0

Calibration as a first-class field.

Every verdict object now returns axes_confidence with per-axis confidence scores and named uncertainty drivers. The driver list is human-readable — "single-LLM derivation (no cross-model corroboration)" — so a downstream caller can show users why the score is what it is, not just what the score is.

Also new: verdict_volatility.would_change_at_score and verdict_volatility.would_change_to. If a verdict is one point from flipping, the response now tells you so. This means an agent can ask "is this borderline?" without running a second call.

Non-breaking. Existing clients ignore the new fields silently.

v2.1.4

USPTO TM corpus refresh restored on a 6-hour SLA.

The USPTO TRTDXFAP daily-delta ingest is now firing on a 6-hour rolling SLA via the new POST /v1/cron/uspto-delta internal endpoint. Verdicts issued after 2026-06-05 06:00 UTC carry the refreshed corpus version under corpus_versions.uspto in the receipt drawer.

Previously the corpus was stale by 144+ hours due to a dead scheduled-task path. The new architecture (board memo 2026-06-02) hardens against silent ETL failure: /healthz now surfaces last_uspto_etl on every response.

benchmark · v0.5

The LLM Hallucination Benchmark v0.5 is live.

1,200 names × 6 frontier models × 6 conditions × 3 repeats = 129,600 generations · 975,192 scored claims. Headline: 97.01% of numeric USPTO citations produced by frontier LLMs while assessing brand names do not resolve to the brand the model was asked about.

The full methodology, the 8.1M-row USPTO reg-and-serial index, and the benchmark dataset are published under CC-BY-4.0 at /research/benchmark. Companion abstract on arXiv pending.

v2.1.0

Seven generator routers will 410 GONE on 2026-07-15.

The following routes are deprecated and will return 410 GONE (not 404) starting 2026-07-15:

  • /v1/generate · /v2/generate · /v2/generate_name
  • /v3/generate · /v3/suggest · /v3/suggest_names
  • /v3/forge/suggest

Reason: Etymolt does not generate names. We verify them. The generator-shape endpoints predate the 2026-05-20 board ruling that locked the category to fact-checking. The 410 response carries a Sunset header pointing to /v1/verify — the right surface for downstream callers.

Six-week notice. If you depend on a deprecated route, email tariq@etymolt.com and we'll discuss your migration.

v2.0.0

Public launch — Etymolt is live.

api.etymolt.com is live, signed, and serving. The MCP server @etymolt/mcp-server is published on npm with OIDC provenance. The seven framework adapters (LangChain, LlamaIndex, CrewAI, LangGraph, OpenAI Agents, Mastra, Vercel AI SDK) ship with cross-adapter disclaimer parity.

Breaking from v1.x: the verdict response shape changed from the legacy generator format to the EVP/1 protocol shape. See /docs/verdict-protocol for the migration guide. The legacy /v1/check endpoint stays canonical; /v3/check now 308-redirects to it.

First five anonymous verdicts free per IP. Ten more on signup. After that, $1 per verdict — no subscription.

v1.9.2

Custom GPT — Brand Name Verifier by Etymolt — is live on ChatGPT.

The Custom GPT calls /v1/verify via the OpenAPI Action and returns the full verdict in chat. Free trial without an API key. Five-axis verdict, signed permalink, every flag traced to a record number.

Link: chatgpt.com/g/.../brand-name-verifier

v1.8.0

Ed25519 signatures on every verdict.

Every verdict now carries an Ed25519 signature over the SHA-256 payload digest. The signing key is etymolt-1779085662, public-key-pinned at /.well-known/evp-keys.json. Key rotation policy: 90-day cadence, 7-day overlap window, both keys signing during overlap.

Downstream agents can verify client-side without a roundtrip. Trust without a roundtrip.