SOTA naming methodology — research findings + pipeline v2 (2026-09-16)

Why the 9-register hunt kept missing the zeitgeist, per the web sweep. Sources below; method retro at the end. This file is the upgrade brief for foundation.json.

The diagnosis (skill vs model vs instructions)

Factor	Weight	Evidence
Method (skill)	~70%	The pipeline used LLMs as GENERATORS. Every 2026 source says the same: that produces “linguistic averaging” — consensus names (Flux, Pulse, Axera, and our ondel/nimbel/avel lane) nobody hates and nobody remembers (Tanj 2026-03; NameStormers “name slop” 2026-01; River+Wolf’s ChatGPT test 2024/26). Lexicon (Placek — Swiffer, BlackBerry, Sonos, Azure, Dasani, Pentium) uses AI only as a PROFILER: 107 linguists + 7 in-house language models score names; humans invent from morphemes (Focus Lab podcast 2026-01-21).
Brief (instructions)	~20%	River+Wolf measured it: a vague prompt → on-the-nose descriptive mush; adding TONALITY + audience + POV dramatically improved output. Our brief had feelings (“floating”, “nextgen”) but no taste anchors (brands the owner loves/hates), no POV decision, no behavior framing. His “onair” instinct is the first real anchor — it names a REGISTER (broadcast/always-live), not just a sound.
Model (GLM 5.3 flash)	~10%	A weaker zeitgeist prior hurts at the margins, but the failure is structural, not model-bound: River+Wolf got the same mush from ChatGPT; Tanj calls it a property of the technique. Model choice matters for PROFILING depth and taste, not for escaping the mush — method does.

SOTA findings, per source

1. Tanj (naming studio, NYC) — 2026-03-05

“AI-generated names all sound the same” = linguistic averaging: models head to the statistical center of “good name” (short, punchy, technical, vowel/x ending). Consensus names. “Great names don’t come from the middle. They come from a decision. A point of view… a name lands because it breaks a rule rather than follows one.”

2. Lexicon Branding × Focus Lab — 2026-01-21 (David Placek, 40yr, the SOTA authority)

• 3 steps: Identify → Invent → Implement.
• Identify = a loose FRAMEWORK (not tight objectives), built on BEHAVIOR, bi-directional: how the company behaves toward the market; how customers behave/feel wearing the name. Name for the future, not for now. The name’s job = “asymmetric advantage”.
• Invent = small teams + linguists extracting MORPHEMES (prefixes/suffixes/infixes) + sound symbolism. 1-2 intense creative cycles. Evaluate on processing fluency. Names presented as TOOLS (“don’t reject anything today”). Quantity leads to quality — “you gotta do more than 50 names.”
• Trademark reality: US class 9 went 23k (1980s) → 2M+ marks; ~9 months minimum per filing; clearance starts EARLY (3-year horizon), licensing/negotiation is a legitimate path. “.com is nice to have — URLs are like area codes.” (validates name.tld≤3 + qualifier .coms)
• Implementation includes performance profiles (linguists profile the name in each language) and an assets/liabilities briefing for the identity firm.
• Founder tip: write the employee announcement for the new name FIRST — imagine the end, work backwards. (We have the positioning statement = the equivalent done.)
• Rename legitimacy: business model changed + an emotional story for the customer. the product passes.

3. NameStormers — 2026-01-08 (“name slop”)

The filter before committing: trademark quick-pass → domain/handles → PHONETIC TEST (say it twice fast, on a call; mishear = kill — this is exactly how ONDIN died) → 3-second category signal → 15-second story test (can you tell why this name matters; “it sounded cool” = fail) → AI = stimulus only, human decides.

4. River + Wolf (Margaret Wolfson — #1 NYC on Clutch, Wharton lecturer) — 2024/2026

• Generation inputs of the pros: bookstores, magazines, product labels, mathematics, music, fashion, poetry, songs, arcane symbol books, foreign languages, etymologies, VC portfolios “to get a sense of the zeitgeist”.
• Lateral (non-linear) paths beat direct description. Suggestive/arbitrary/fanciful > descriptive (also the trademark-strength order).
• Prompt-shape effect measured: tonality + audience in the prompt = visibly better output.
• Foreign-equivalents doctrine: a French/ES word familiar to US consumers gets translated by examiners (sans = without) — our onde/onda root reads through.

5. Crunchbase era map — 2025-05-23 (the zeitgeist, dated)

2016-17 misspellings → 2017-18 animals → 2018-19 fake-real coinages → 2019-20 literal descriptions → 2021-22 -ly/-ify suffixes → 2023-26 the “boring/serious” era: real English words with meaning (Anthropic, Perplexity, Linear, Sierra, Clay), serious composites (OpenAI, CoreWeave, Safe Superintelligence), iconic short oddballs (Suno, Groq, Arc, Attio), first names (Theo), animals (Project Orca). Zero-anchor coinage mush is the AI-consensus look — the opposite of this era’s winners.

6. The science layer — sound symbolism (replaces vibe-registers)

• Klink, “A Sound Idea: Phonetic Effects of Brand Names on Consumer Judgments” (J. Consumer Research): vowel/consonant classes carry measurable attribute associations (front vowels /i/ = small/fast/bright; back vowels = large/heavy; fricatives = fast/smooth; plosives = solid/rugged).
• Phonesthemes paper (J. Consumer Psychology, 2024): non-word brand names can be deliberately imbued with meaning via phonesthemes — the evidence-based way to coin (vs zero-anchor mush).
• Implication: the position “runs itself / weightless / always-live” maps to a PHONEME PROFILE (fluids+fricatives l/s/v/f, open bright vowels, sonorants), and every coinage must carry a morpheme story (onde-, -air, -va→no) — the feeling must be citable.

Pipeline v2 (the upgrade)

1. BRIEF first (gates generation): taste anchors from the owner — 5 brands loved, 5 hated, the names he’d have bought if free; enemies-not-to-sound-like (have: Zenoti/Planity/Fresha); the behavior framing (have: “my business runs itself — I only do the craft”); the POV sentence the name must break-a-rule TOWARD (“onair” instinct = the always-live register).
2. Invent from morphemes + phonesthemes, never from vibe-words alone:
   • Build the MORPHEME BANK for the position (live/air/broadcast: on-, -air, signal, cue, wavelength, seg-, live-, -cast; smooth/auto: gliss-, liss-, flu-, -flow; FR wave: onde-).
   • Lateral discipline paths (broadcast, music tempo, sailing, timekeeping).
   • Phoneme profile per candidate + one-line morpheme story REQUIRED before screening.
   • Quantity: ≥50/cycle, 2 cycles, “don’t reject anything today” mindset with the owner.
3. AI as PROFILER (inverted!): 0h cold-read (have) + NEW gates: phonetic/stress-ambiguity (“say it twice fast”), 3-second category signal, 15-second story test, zeitgeist check (“does it sound like 2023-26 winners or like Axera?”).
4. Zeitgeist corpus, maintained: Crunchbase-era winners list + VC portfolio names (Y Combinator batches, a16z) as the “what good sounds like NOW” reference.
5. Ownability early: RDAP .ai (have) + USPTO/EUIPO knockout before attachment; qualifier .coms fine (Lexicon: URLs are area codes).

Retro (research-channels)

GOOD: four primary practitioner sources + era data + science layer in one session. BAD: dfs rate-limited mid-sweep; guessed URLs cost 3 failed fetches. UGLY: dfs message-boards lane returned SEO mirror-spam for the commercial naming topic (known boundary, reconfirmed) — Reddit stays snippet-level only (Google-index ceiling).

Sources

• tanj.co/blog/why-ai-generated-names-all-sound-the-same/ (2026-03-05)
• focuslab.agency/blog/brand-naming-in-2026-ai-trademarks-emerging-trends (Placek, 2026-01-21)
• namestormers.com/how-to-avoid-name-slop-why-ai-generated-names-fall-short-what-to-do-about-it/ (2026-01-08)
• riverandwolf.com/brand-naming-is-tricky-and-generative-ai-cant-do-it-yet/ (2024-12-15, upd 2026-08)
• news.crunchbase.com/startups/historic-naming-trends-weird-boring-misspelled/ (2025-05-23)
• Klink (JCR) + JCPY 2024 phonesthemes — see SERP-captured URLs in naming-log §15