SOTA naming — delta brief (2026-09-30)

Scope: what has CHANGED or is MISSING since the doctrine’s last hardening. The doctrine is already current through 2026-09-16 (SOTA_UPGRADES_V2 / FOUNDER_TASTE_ANCHORS) with a 2026-09-30 graveyard entry, so this brief is a delta against that, not a restatement. Every claim carries a URL + source date. Claims I could not verify the specifics of are marked [UNVERIFIED].

1. Name generation SOTA

Nothing in pipeline_tools has died — all five named tools are alive in 2026. But the toolset is missing two 2026 shifts: “registry-confirmed” AI generators, and agent-native registrar APIs.

• LeanDomainSearch — still alive. Loads as a wp.com-hosted page; the /search/ filters (Starts/Ends with term, Popularity/Length/Alphabetical) are still present. https://leandomainsearch.com/ (fetched 2026-09-30, live page). → no action (keep it in the list).
• Namelix — still alive, now runs a “custom LLM,” no longer GPT. Homepage: “Generate more original ideas than ChatGPT using our custom LLM”; also ships a free “Namelix Domains — all domains free to register” sibling service and Brandmark.io logo tie-in. https://namelix.com/ (fetched 2026-09-30, live page). → no action, but the “custom LLM” wording confirms Namelix is ideation-only (matches our “never clearance” rule, unchanged).
• ExpiredDomains.net — still alive, bigger. Tracks 676 TLDs / ~770M total domains / ~584M deleted; the dev feed shows active registry work as late as 2026-06-18. https://expireddomains.net/ + https://www.expireddomains.net/feed/dev (search-verified, feed dated 2026-06-18). → no action.
• Spamzilla — still alive, now has a FREE tier. Pricing page shows FREE $0/mo — Review Only <PetName>25 Domains alongside paid “Full Access Millions of Expired Domains.” https://www.spamzilla.io/pricing (search-verified, live page). → minor: our “vet expired” step can now run free for small batches — note the free tier in pipeline_tools.
• DomainsGPT — still alive. Now presented as a “domain search engine … powered by the Domai.io API,” not just a GPT wrapper; output pairs a GoDaddy Valuation with a live registration price side by side. https://www.domainsgpt.ai/ (fetched 2026-09-30, live page). → copy the pattern: show appraisal + reg-price side by side in our own domain screen.
• [NEW] NameBuddy.ai — 2026 entrant aimed squarely at our pain point: “Registry-checked • No affiliate links • Never ‘premium’ prices … AI invents it. The registry confirms it.” https://namebuddy.ai/ (fetched 2026-09-30, live page). → add to pipeline_tools — it is the first tool whose pitch is “no DNS false-positives,” i.e. the exact failure the doctrine flags.
• [NEW] registrars are becoming agent/API-native. Porkbun publishes an official MCP server (npx -y @porkbunllc/mcp-server), an llms.txt, a sandbox API, and dryRun:true on every billable op including domain checkDomain (availability + price in one call, idempotency-keyed). https://porkbun.com/llms-full.txt (fetched 2026-09-30). → high-value: replace the MCP DNS checker with registry-grade availability calls (Porkbun checkDomain/dryRun) — directly fixes DO_NOT_REVERT #5’s “MCP false-positives.”
• [NEW] instantdomainsearch now bundles an “Expired domains” flow + an AI generator (see §6). https://instantdomainsearch.com/ (fetched 2026-09-30). → no immediate action; watchlist.

2. Sound symbolism / psycholinguistics

The doctrine’s phoneme-profile layer (Klink vowel/consonant classes + the 2024 phonestheme paper) is not contradicted by any 2026 finding I could verify. Two 2026/2025 items extend it, and one cross-cultural study is a flag, not a contradiction.

• [2026] “Neural processing of sound-symbolic associations in product branding.” APA PsycArticles, record 2026-57229-002. Paywalled (HTTP 403); I verified the title/record and year only — specific findings [UNVERIFIED]. If it holds up it would give the phoneme-profile gate a neural (EEG/fMRI) underpinning rather than a behavioral one. https://psycnet.apa.org/record/2026-57229-002 (record, 2026). → no rule change yet; re-check the abstract before citing it as neural evidence.
• [2025] “Beyond Words: A Cross-Cultural Comparative Study on Brand Naming Strategies and Sound Symbolism.” Zulfugarova et al., J. International Consumer Marketing 38(4), 2025. DOI 10.1080/08961530.2025.2594791. Paywalled — findings [UNVERIFIED]. https://www.tandfonline.com/doi/full/10.1080/08961530.2025.2594791 → flag, don’t change: Klink’s direction (front-vowel = small/fast/bright, back = large/heavy) does generalize across English/French/Chinese per the earlier “Creating Global Brand Names” lit (ResearchGate synthesis), but the doctrine should treat its phoneme rules as Western-anchored until this Türkiye-based study is read — which is exactly what GATE “native reads in every launch language” already enforces. https://www.researchgate.net/publication/271822468_Creating_Global_Brand_Names_The_Use_of_Sound_Symbolism
• [background, still consensus] Phonestheme imbuing (“Snyre for your nasal congestion”) — the J. Consumer Psychology 2024 non-word-meaning paper the doctrine already cites; Monash-hosted OA copy confirms it remains the citable reference. https://researchmgt.monash.edu (search-verified). → no action (already in doctrine).

3. Domain market reality

Biggest block for us: the verification path for .ai/.io changed, .com got pricier, and a wave of new strings is incoming.

• [.ai / .io RDAP situation — the one the doctrine must absorb.] IANA delegation for .ai now list RDAP server https://rdap.identitydigital.services/rdap/ (record updated 2025-02-11), and Identity Digital’s own whois banner at whois.nic.ai says: “The Identity Digital Whois service may no longer support all TLDs operated by Identity Digital … please use another publicly accessible RDAP tool.” https://www.iana.org/domains/root/db/ai.html + http://whois.nic.ai/?q=resumed.ai (fetched / search-verified, 2026-09-30). .io is also under Identity Digital (OpenSRS status incident lists .me .info .io .ai .group .ltd as “Identity Digital Registry Connection,” 2026-08-05). https://isdown.app/status/opensrs/incidents/634312-identity-digital-registry-connection-me-info-io-ai → update the verification gate: for .ai/.io (and any Identity Digital TLD), stop relying on port-43 whois; query RDAP at rdap.identitydigital.services/rdap/domain/<name>. This matters because DO_NOT_REVERT #5 already proves whois false-positives on non-.com are catastrophic.
• Verisign .com wholesale fee +7%: $10.26 → $10.97 effective 2026-11-01 (first increase allowed since the Feb-2024 notice). Primary: Verisign Q1 2026 results. https://investor.verisign.com/news-releases/news-release-details/verisign-reports-first-quarter-2026-results/ (fetched, 2026-04-23). → minor: re-run the defensive-set math — “grab a typo or two ~$10 each” becomes ~$11 and renewals creep; not material enough to change policy, but update the $ figure.
• .web finally delegated (2026-07-22). Verisign plans to sell .web via channel partners “later this year.” https://investor.verisign.com/news-releases/news-release-details/verisign-announces-delegation-web (2026-07-22, search-verified primary). → watchlist: .web is the most “generic-feeling” new string in a decade; when it launches, reconsider the novelty-TLD blacklist’s stance on it.
• ICANN next round is done applying: 1,600+ primary applications (1,100+ replacement strings), window closed 2026-08-12. Primary. https://www.icann.org/en/announcements/details/icann-2026-round-closes-with-more-than-1600-new-gtld-applications-13-08-2026-en (fetched, 2026-08-13). Reveal Day ~9 weeks post-close (≈ October 2026). → add a standing note in TLD_DOCTRINE: a flood of new strings lands 2027-28; also WIPO is now the exclusive LRO/SCO objection provider for the round (source: WIPO 2025 record, below).
• .ai liquidity tripled. 2024 .ai aftermarket ≈ $9.4M → 2025 ≈ $27.1M (189%); .ai has 600,000+ registered names and was fully migrated to Identity Digital 2025-01-15. https://www.identity.digital/newsroom/ai-completes-a-historic-migration-to-the-identity-digital-platform (fetched, 2025-01-15) + https://www.thetimesherald.com/press-release/story/41205/… (Mar 9 2026, secondary). .ai registry price is rising too (EuroDNS .ai 132€ → 140€ effective 2026-03-05). https://webhosting.today/2026/01/31/eurodns-updated-its-prices-for-29-domains (fetched, 2026-01-31). → reinforces, doesn’t change the doctrine’s “.ai is first-class for AI-native buyers” line; only tightens the “buy now, .ai gets pricier” urgency.
• .io is NOT going away in 2026. May-2025 UK/Mauritius treaty; IO code still in ISO 3166-1; any retirement would be a multi-year ICANN phase-out. (Secondary vendor source.) https://bolddomains.com/blog/is-the-io-domain-going-away (fetched, 2026-07-19). → no change to the “Chagos-flagged dev-only” posture — tail risk unchanged, not eliminated.

4. Trademark / legal

The one-hour legal gate still covers the realistic failure modes, but two 2026 updates sharpen it — and one UDRP decision narrows a defense we lean on.

• WIPO UDRP fee schedule + new services (2026-03-09): added expedited case processing (decision in ~1 month; $4,000 complainant fee for 1–5 names / $2,500 respondent) and a reduced early-termination fee ($100 retained for pre-notification withdrawal). Primary. https://www.wipo.int/amc/en/domains/news/2026/news_0005.html (fetched, 2026-03-09). → update LEGAL gate step 5 cost figures (UDRP is no longer flat ~$1,500; add the $4,000 expedited ceiling and note faster timelines cut both ways).
• 2025 = record UDRP year: 6,200+ cases; 80,000+ over 25 years. Primary. https://www.wipo.int/amc/en/domains/news/2026/news_0001.html (fetched, 2026-01-14). Two case notes matter to us directly:
  • D2025-3938 — the “brandable portfolio” defense was narrowed. A respondent arguing a name was a dictionary-brandable portfolio item was told to show the domain “conveys a linguistic or conceptual meaning independent of the complainant’s mark.” This is precisely our aftermarket buy path. → add to LEGAL gate step 5: “brandable” is NOT a standalone UDRP defense; a bought brandable must carry independent linguistic meaning.
  • DAI2025-0060 — choosing .ai as (part of) a complainant’s mark tipped the balance against the respondent. → reinforces that for an AI-native brand, the .ai string itself can become part of the protected mark (registration path, not just domain).
  • D2025-2916 — panels warning on AI-drafted pleadings. Relevant if we ever auto-author UDRP filings: AI “may complement but [is] not a substitute for professional advice.”
• USPTO: backlog inside the register is IMPROVING. “March sets record for Trademarks first action output as filings rise and inventory declines” (bulletin 2026-04-23); FY2027 justification on file. https://content.govdelivery.com/accounts/USPTO/bulletins/4140a7e (search-verified, 2026-04-23). → soften the doctrine’s “register is saturated ~10mo to registration” doom to “saturated but first-action output now at record, inventory declining” (still clear early; pace is moving in our favor).
• Fee ballast, 2026: UK IPO raises trade-mark fees 2026-04-01 (primary) https://www.gov.uk/government/publications/intellectual-property-office-new-fees-from-1-april-2026/… ; EUIPO office fees do not change in 2026 (secondary) https://www.markencheck.ai/en/blog/eu-trademark-fees-2026 (2026-06-26); Madrid US individual fee change (WIPO Info Notice 21/2026). https://www.wipo.int/documents/d/madrid-system/information-notices-en-2026-madrid_2026_21_e.pdf → minor footnotes only; no gate change.

5. Cold-read probing / LLM sycophancy (most important — you’re about to ship this)

Current protocol (≥3 families × ≥5 runs, third-person framing, mode + dispersion) is directionally right but under-specified on three points the 2026 literature now nails down. Sharma et al. 2023 (arXiv 2310.13548, sycophancy in LLMs) and the Nov-2024 survey (arXiv 2411.15287) remain the baseline; the new work is about judge reliability and independence, which is exactly what a probe score depends on.

• [RUN COUNT] ≥5 runs is too few — target ≥11. “The Coin Flip Judge?” (Yagubyan, arXiv 2606.13685, 2026-04-23): repeated identical evals flip pairwise preferences 13.6% of the time (28% of questions >20% flip, one at 56%); 11 repeated trials are needed for a majority vote to recover the 50-trial reference verdict with 95% probability (15 for high-variance questions); cross-judge agreement is only 76% (κ=0.51). https://arxiv.org/abs/2606.13685 (fetched). → change the probe: ≥11 runs ≥3 families, and report the flip-rate/dispersion, not a single mode.
• [INDEPENDENCE] ≥3 families may not be independent. “A Statistical Framework for Auditing Behavioral Dependence and Induced Bias in LLM Judges” (Kuai et al., COLM 2026, arXiv 2604.07650, 2026-04-08): across 18 LLMs / 6 families, shared pretraining-distillation-alignment induces “latent entanglement” → correlated reasoning and synchronized failures; entangled judges show over-endorsement bias (ρ≈0.51). https://arxiv.org/abs/2604.07650 (fetched). → change the probe: diversify beyond family (instruct-style / no-system-prompt / different decode temp) AND add one non-LLM or human spot-check so 3 “families” aren’t a monoculture.
• [SCORING] Mode+dispersion is not a validity statistic. “Reliability without Validity” (Norman et al., arXiv 2606.19544, 2026-06-17): across 21 judges / 9 providers / 118 runs / 541k judgments, exact-match agreement overstates discriminative ability by a 33–41pp kappa deflation; judges re-rank by up to 14 positions across benchmarks; high test-retest (>0.95) coexists with severe position bias (>0.10). https://arxiv.org/abs/2606.19544 (fetched). Companion “Agreement Metrics for LLM-as-a-Judge” (Rao & Callison-Burch, arXiv 2606.00093, 2026-05-25): protocol choices alone move reported “accuracy” 0.551 → 0.899 without changing a single verdict. https://arxiv.org/abs/2606.00093 (fetched). → change the probe: pin a chance-corrected score (Cohen’s κ / ICC against the category spread), and report the protocol (abstentions, pooling, scale) so the number is reconstructible.
• [SELF/FAIM] families favor their own. “Play Favorites” (Spiliopoulou et al., arXiv 2508.06709, 2025-08-08): 9 judges on >5,000 pairs — GPT-4o and Claude 3.5 Sonnet systematically rate their own and same-family outputs higher. https://arxiv.org/abs/2508.06709 (fetched). → change the probe: if any candidate is ever AI-generated, score it with families disjoint from the generator family; keep judge families recorded per run.
• [FRAMING] third-person framing is empirically justified. “Sycophancy Negatively Affects LLM-as-a-Judge in Conflict Evaluation” (Farzi, Dietz, Carton, GEM 2026, 2026-07): replacing a speaker with first-person “Me” biases blame/responsibility judgments toward the narrator. https://aclanthology.org/2026.gem-main.45/ (fetched). “Judging the Judges” (ACL 2025, IJCNLP) confirms position bias in judge prompts. https://aclanthology.org/2025.ijcnlp-long.18/ → confirm, don’t change: keep third-person framing; additionally randomize any ordering.

Net protocol delta: ≥3 families is only meaningful if the families are disentangled (mix instruct styles/decoders + one human spot-check), ≥11 runs (report dispersion not a singleton mode), score is chance-corrected (κ/ICC), judges never share a family with any generator, and framing stays third-person with randomized ordering. Each of the five now has a 2026 citation to lean on.

6. Competitor IA and UX

• instantdomainsearch.com (fetched 2026-09-30) — now leads with a tools nav, not just the type-head search: Domain extensions · Domain generator (AI) · Premium domains · Expired domains (marked “New”) · Bulk domain search, under a “Finding domains” heading. The single-search-tool has sprouted a full toolbox. → copy the pattern: search-first, but surface “generator / premium / expired” as first-class tabs. Avoid: their monolithic 800+ extension list.
• Namelix (fetched 2026-09-30) — three-step value prop (AI naming engine → filter by length/keyword/extension → save & it learns), zero nagivation chrome, one call-to-action. → copy: lead with the job-to-be-done steps, not a feature list. Avoid: “custom LLM” opacity (we should state our scoring method).
• LeanDomainSearch (fetched 2026-09-30) — still the same minimal “type a word, get hundreds of pairings” flow with Starts-with/Ends-with filters. Dating itself. → avoid its dated look, keep its speed-to-results.
• DomainsGPT (fetched 2026-09-30) — the one IA idea worth stealing: every suggestion renders a third-party appraisal (“GoDaddy Valuation: $1,183”) beside the registration price (“$37.38”). → copy exactly this on our appraisal screen: independent valuation + live cost, side by side.
• Porkbun (fetched 2026-09-30) — ships an official MCP server, llms.txt, sandbox + dryRun API. Not a UI pattern but a strategy pattern: registrars now court AI agents as first-class buyers. → adopt: a naming tool that can verify/register via a registry-grade API beats any MCP DNS checker; this is our wedge.
• GoDaddy — bundles Airo® (AI website/logo/name builder) free with every domain purchase; “Airo for WordPress” shipped 2026-05-11. https://aboutus.godaddy.net/newsroom/news-releases/press-release-details/2026/GoDaddys-Airo-for-WordPress-Delivers-AI-that-Builds-Grows-and-Continuously-Improves-Websites/default.aspx (2026-05-11). → match the bundle if we stay in the registrar orbit; otherwise stay the independent appraisal/name source.
• Dan.com — gone; listings migrated to Afternic (GoDaddy). https://domainsdealer.com/blog/what-happened-to-dan-com (2026, secondary) + https://www.afternic.com/ (search-verified, live). → correct the competitor list: Dan.com is dead, Afternic is the venue (we already cite Afternic in the aftermarket playbook — now it’s the only GoDaddy marketplace path).

UNVERIFIED (specifics not confirmable from the cited page)

• “Neural processing of sound-symbolic associations in product branding” — 2026, but abstract/findings paywalled; only title/year verified.
• “Beyond Words: A Cross-Cultural Comparative Study…” (JICM 38(4), 2025) — DOI/title/authors verified; findings paywalled.
• The whois.nic.ai deprecation banner is live/undated (verified current as of 2026-09-30, not dated).
• GoDaddy Airo full feature set / whether it generates names vs only builds sites — verified only that Airo exists and “Airo for WordPress” shipped 2026-05-11.

Highest-value deltas

Ranked by impact on the product:

1. Cold-read probe is under-powered and possibly non-independent — tighten before you ship. (Coin Flip Judge: 11 trials needed, 13.6% flip; COLM 2026: 3 families can be entanglement, not independence; Play Favorites: family self-bias; Reliability-w/o-Validity: mode% overstates validity 33–41pp.) Action: ship ≥11 runs × genuinely disentangled judges × chance-corrected (κ/ICC) scoring × disjoint generator/judge families + one human spot-check.
2. The .ai/.io verification path just moved to RDAP (Identity Digital), and whois is being sunset. Action: swap the availability gate for Identity Digital TLDs to rdap.identitydigital.services/rdap/domain/<name>, and add Porkbun’s checkDomain/dryRun as the registry-grade MCP replacement.
3. The “brandable” aftermarket-defense just narrowed (WIPO D2025-3938) and UDRP got an expedited $4,000 track. Action: add to the legal gate — a bought brandable must carry independent linguistic meaning or it is cybersquatting on a mark; budget $4,000 for the fast lane.
4. Verisign .com +7% to $10.97 (2026-11-01) and .web went live (2026-07-22). Action: bump the defensive-variant cost numbers, and add .web to the TLD_DOCTRINE watchlist for the novelty-blacklist review.
5. ICANN 2026 round closed with 1,600+ new strings (Reveal Day ≈ Oct 2026). Action: add a standing TLD_DOCTRINE note to re-sweep at Reveal Day for strings that collide with client names (WIPO is the exclusive LRO/SCO provider).
6. Registrars are going AI-agent-native (Porkbun MCP + dryRun, GoDaddy Airo). Action: make our pipeline call a registry-grade API instead of the false-positive-prone MCP DNS check — the single highest-leverage tooling upgrade available.
7. USPTO trademark first-action output is at an all-time record and inventory is declining (2026-04-23). Action: soften the legal gate’s “~10-month saturated register” pessimism to “saturated but accelerating” — keep clearing early, but signal the pace is moving in our favor.