Generation methods

THE #1 METHOD for a free + great + .com name in 2026 (added 2026-07-06, research-backed). His own exemplars are BLENDS of two real words, NOT single dictionary words: ByteDance = byte(tech)+dance(art/social); Pinterest = pin+interest; Instagram = instant+telegram; Airbnb = air+bnb. Single real words are ALL taken (registered decades ago, $10k-$10M aftermarket); the novel BLEND is FREE precisely because nobody paired those two words. THE KEY = DISTANT LEAP: the two words must come from FAR-APART domains that resolve into the business (tech↔dance, not trade↔wind). OBVIOUS same-field pairs are taken; the SURPRISE of the leap is what keeps the .com free AND makes it click. METHOD: (a) name the activity anchor; (b) brainstorm words from a DIFFERENT domain — nature, art, music, animals, physics, motion, weather, warm-myth — that capture the SOUL/result; (c) blend; (d) bulk .com-check (RDAP); (e) HUMAN-curate for the click (only a person judges 'does the leap resolve AND spark'). Blends are also 100% trademark-able. This is ON-TASTE (it IS the ByteDance move he named), .com-friendly (novel=free), and the ONLY path that satisfies all his constraints at once. NOT soft-consonant coinage (Register A) and NOT a smoothie-mash of sounds — two WHOLE real words, cleanly joined.

For Register-C WINNER names (house/holding/flagship) — HIS named method: generate like naming a RACEHORSE, a destiny-claim to be SHOUTED by the announcer at the finish (full doctrine = WINNER_NAME_DOCTRINE_2026_07_09). Recipe: (a) build an APEX wordlist (money/space/empire/king/dark/gold-class ambition-domain words) × (b) a PERCUSSION wordlist (shot/X/dance/horse/strike-class punchy action) with the pair INTERACTING (verb-grammar, never noun-concatenation); (c) ANNOUNCER-TEST each survivor aloud ('AND HERE COMES ___!'); (d) tri-civilization echo check (China 黑马/长期主义 · MENA rahma/Al-'Adiyat/sabr · West idiom); (e) WebSearch collision check (the step that killed evertide/turnstone/ironflow); (f) GRIN test — does saying it produce a smirk? Study the actual turf registry for patterns: Secretariat, Man o' War, Empire Maker, Bold Ruler, Sunday Silence, Flightline, Seattle Slew.

(Lexicon — Swiffer/BlackBerry/Sonos/Vercel) Generate 1,000+ candidates, shortlist 50-100. A 20-name brainstorm is structurally under-powered. YC data: 87% choose brandable over keyword; median 8 chars (6-10 sweet spot); 2026 momentum = unexpected compound pairings (specific enough to feel real, broad enough to pivot). Lexicon: compounds are '1+1=3' (PowerBook/BlackBerry) — richer associations beat shorter-but-empty coinages.

(Lexicon anti-anchoring)

ALSO THE #1 ANTI-SLOP WEAPON (2026-08-27) — the documented fix for BOTH human groupthink AND AI linguistic-averaging is identical: DISGUISE / DIVERGE THE BRIEF. Run parallel lanes: one with the real brief, one 'naming a competitor', one naming an UNRELATED category (Windsurf came from the unrelated team). Lexicon deliberately feeds its human teams disguised briefs to protect originality; the same trick forces an LLM off its statistical centre. = the divide-and-conquer agent fan-out, formalized. Pair with GATE 0g (slop check) on the output.

Define before generating: the Win / what we have / what we need / what the name must SAY — name the desired behavior/experience, not the product.

(2026-08-27) THE OPERATIONAL FIRST-READ TEST (the executable half of GATE 0h; run on every finalist BEFORE the collision sweep — it is cheaper and kills faster). PROTOCOL: (1) zero-context probes, one question per FRESH session (no memory/personalization), ≥3 model FAMILIES (not 3 runs of one — bias is family-correlated), ≥5 runs each at default temperature (the SPREAD is the signal); the FLOOR is 12 readings — below 11 a majority vote does not recover the reference verdict (sota-methodology §5), and a 9-reading pass does not rank like a 15-reading one (rho -0.16, 2026-11-04); (2) third-person framing always — 'a company is called X: what do they sell? what industry? first-guess?' — NEVER 'I'm considering naming my company X' (sycophancy caves the test, Sharma 2023); (3) add the shortlist probe: 'best <true category> providers for <buyer>?' — does the name even belong on that list's register? Require a machine line ('BELONGS: yes' / 'BELONGS: no') and read THAT, never the prose: the model names the name in every reply, so a substring test scores 'No, X would not belong' as a fit (measured 2026-11-04). A reply with no line is UNKNOWN, never a no; (4) score by MODE + DISPERSION, never a rating scale (LLM numeric ratings trend random — NeurIPS 2024): ≥60% converge on a wrong/negative category = HARD FAIL; scattered <25% clusters = pass. CALIBRATION: the invoicing-read and the finance-read must both FAIL under this rule — if a protocol change lets them pass, the protocol is broken. SCOPE LIMIT: comprehension screening ONLY — never LLM preference-ranking between survivors (evidence: synthetic respondents flatten variance + skew positive; human free-association or a real panel decides preference). Subagent-friendly: N context-free subagents each answering the probe = the in-harness implementation.

Generate→bulk-check replaces hand-guessing: LeanDomainSearch (word-pairing, .com-only, thousands/sec) + Namelix (ideation ONLY — outputs mostly already registered) + Porkbun/GoDaddy BULK check (paste wordlists) + ExpiredDomains.net (daily droplists, filterable) + Spamzilla (vet expired for spam history). Always vet expired names: Wayback history, backlinks, trademark conflicts. AI generators: volume only, NEVER clearance (they hallucinate availability and regurgitate near-existing marks).

2026 AI generators for the VOLUME half: DomainsGPT (OpenAI-based brandable/blend generator), Instant Domain Search generator (free, live availability + registrar prices), Namelix. Best use = feed them the DISTANT-LEAP-BLEND brief, bulk-check the output, then HUMAN-curate for the click + clearance (AI can't judge whether the leap resolves).

THE SECOND BREAKTHROUGH (2026-07-09, HK-holding hunt — a full live session of his vetoes distilled). His THIRD exemplar surfaced: MoneyShot (joining ByteDance + SpaceX). The unlocking decomposition (validated by him: 'well thus… really the key'): winner names = un mot APEX (money, space) + une PERCUSSION (shot, X, dance). Le nom lui-même projette LA VICTOIRE — pas la stratégie, pas la sagesse: LA DOMINATION AVEC UN SOURIRE. And the METHOD he named himself: 'you got to learn how to find winner names like they find horse names for the horse race.' Both are now LAW for Register-C house/holding/flagship names.

[APEX word — ambition/dominance domain: money, space, empire, king, dark(mystery), gold…] × [PERCUSSION — short punchy action/impact: shot, X, dance, horse, strike…]. The name must PROJECT VICTORY ITSELF. Wise fails, poetic-serious fails ('not warm like ByteDance or SpaceX or MoneyShot'), generic fails (Keystone: 'the most generic ever, no twist when presenting').

Racehorse names are DESTINY-CLAIMS written to be SHOUTED by the track announcer at the finish line (Secretariat! Man o' War! Empire Maker! Bold Ruler! Sunday Silence! Flightline!). THE ANNOUNCER TEST: shout 'AND HERE COMES ___!' at the moment of victory — does it sound like a winner crossing the line? If it sounds like a strategy, a place, a mood, or an apology, it fails. Turf craft to copy: 1-3 punchy words ≤18 chars, attitude + poetry + a wink, unexpected collisions that feel inevitable.

The twist = one part ACTING ON the other (byte DANCES · rain is MADE · check MATE). Noun+noun = concatenation = dead (the noun+noun kill). a near-yes of that session — VALIDATED the grammar and died only on the verb's semantics ('leap' = frog/gadget).

'Warm' in his mouth = THE GRIN: saying the name produces a smirk/lean-in (MoneyShot's cheek, ByteDance's joy, SpaceX's audacity). NOT gentle/poetic. CRITICAL CALIBRATION of the sound-veto: MILD TRANSGRESSION IS GOOD (MoneyShot's double-entendre = edge = menace tranquille); what kills is WEAK / RIDICULOUS / SCAMMY echoes — anatomy (a word that reads as a body part), apology (one that reads as an apology), a finance/trading/gambling smell, gadget (the 'leap' word reads as a frog).

His rule verbatim: 'a Chinese will give more importance to a ROAD or something HEAVY PATIENT STRONG than a leap like a frog he'll find gadgety.' Weight-words (road, horse, iron, stone, mountain, tide) > cleverness-words (leap, spark, dash). The dynamism must be the HEAVY kind — a racehorse, a tide, a road — never the light kind.

The house is WORLDWIDE — China+West+MENA-friendly, specific to NONE (this pivot killed the whole silk/jade lane mid-session). Check every finalist's echo in all three: China (黑马 dark horse = beloved business compliment · 长期主义 · 及时雨), MENA (horses sacred/Al-'Adiyat · rain = rahma · sabr · iron = Surah Al-Hadid), West (the idiom read). Positive in all three or dead.

HIS OWN CALL 2026-07-09 ('ok lets find even without dotcom'): a Companies-Registry LEGAL name is NOT a domain purchase → the HK-Ltd legal name may proceed without its .com. The RANKED_TLD_POLICY (2026-08-27; .com default, .ai for an AI-native brand, buyer-matched) governs any domain/brand web presence (buy path if taken — $3-15k two-word). ALSO learned: the bland-placeholder DECOUPLE failed in practice — he rejected Keystone because he must SAY the legal name aloud; even the legal name must pass the announcer/grin bar. AND: never suffix 'Group/Holdings' on the HK Ltd (passive-holding flag → bank EDD, hurts HSBC Business Express fast-track + GEP operating narrative).

Lead candidate produced BY the method — an announcer-native turf term; 黑马 = admiring Chinese business idiom; horses sacred in MENA (Al-'Adiyat); DARK = his Bruce-Wayne/menace-tranquille register; house-of-brands = THE STABLE ('we breed dark horses'); dark+horse = A1 ESL + radio-perfect. Honest flags: Dark Horse Comics (US publisher) + wine + song = known namespace, but nobody owns it in business/holding. STATUS: PENDING his verdict — NOT locked. Soft-rejected runner-up: Rainmaker (MoneyShot-in-a-suit, 及时雨, rain=rahma — sub-winner-grade for him; cloud-seeding startup collision).

(2026-07-10 — the missing law) The percussion/twist must be TRUE TO WHAT THE COMPANY DOES, not decorative. ByteDance's 'dance' has percussion value BECAUSE they make TikTok dance videos — the twist is EARNED by the activity. A pairing where the twist says nothing about him FAILS even if it scans perfectly (runevolt — 'volt' is decorative = dead; apex×random-digital = dead). AND: NO CHILDISH ANIMAL NAMES for a serious house of brands (lion/tiger/wolf/raven/falcon/cobra/viper all RETIRED 2026-07-10 — 'we don't need a childish animal name'). The tool (greathunt) NETS availability; whether the twist RESONATES with what he does is a HUMAN call — aim it at the soul (next key), then curate.

(2026-07-10) The TWO anchors the earned twist must encode: (a) a HOUSE OF BRANDS building MANY products (the portfolio’s products…) = a forge/studio/factory of ventures, one parent many brands; (b) THE underdog-rise story — 'starting from nothing to everything', the relentless underdog→empire (he loved it; only 'dark'=shady-in-business + animal-fatigue killed it, NOT the underdog-racing soul — that soul is still the target). The idiom-words that nail it are ALL taken by active software cos: wildcard (YC AI-agent startup), moonshot = Kimi's $20B Chinese AI parent Moonshot AI (impossible in his region), longshot/underdog/upstart/groundswell (.com all owned). So generation aims at [builds-many ∪ underdog-rise], non-animal, and the winner is a human call on earned-twist resonance — NOT the tool's mechanical score.

Pipeline v3 - THE CONSTRUCTION DOCTRINE (founder demand: 'improve the skill again, make a breakthrough'). The structural insight after 12 rounds: every prior cycle mined one of two quadrants - (1) REAL WORDS (high density, all taken: 24/25 on .ai) or (2) LOW-DENSITY COINAGES (free but mush: tidezo/dawnva = zero phonological neighborhood). The unmined quadrant: LEXICALLY EMPTY + PHONOLOGICALLY DENSE - constructed words that FEEL like real words (high neighborhood density: plan/sanity/plain around planity; glissando around lissando) while being legally empty (trademarkable, domains free). EVIDENCE: Vitevitch & Donoso 2012 'Phonotactic Probability of Brand Names: I'd buy that!' (PMC3289729) - brand names with common phoneme sequences raise PURCHASE INTENT; phonological neighborhood density (PND) is a 20-yr psycholinguistics program (Vitevitch-Luce review) - high-PND nonwords are perceived, remembered and TRUSTED like real words; BrandScout 2026-03 analysis of 4,000 brands: coined names following natural phonotactics outperform real-word names 2x on trademark-ability and +31% on 48h recall. MECHANISM (naming/tools/construct.py): morpheme matrix (product-real roots x quality tails from the anchor morphology: -oti/-ity/-eau/-elle/-ado/-ora) -> hard filters (not in dictionary, 4-8 letters, 2-3 syllables, no banned -ia/-va/-ea tails, vowel rhythm, slurs) -> SCORE = char-bigram log-prob from 10k common words (phonotactic naturalness) + log(neighborhood density) where density = dictionary words within edit distance 2 (SymSpell deletion-variant method, 370k dict, <10s). RESULT round 13: 18/26 constructed candidates FREE on .ai (vs 24/25 TAKEN for real words) - the domain wall is GONE in this quadrant. LLMs remain PROFILERS (0h probe); the engine generates, the panel discriminates. Source note: wordlists dwyl/english-words + google-10000-english (curl raw.githubusercontent, see script header).

THE RESEARCH SWEEP BEFORE SHIPPING THE PROBE (2026-09-30; full brief + URLs: ~/work/neo/naming/context/sota-methodology-2026-09.md, ~20 verified sources). Four durable changes, each with its citation. (1) COLD-READ PROTOCOL v3 — SUPERSEDES the '>=3 families x >=5 runs' half of GENERATION_METHODS.cold_read_probe. RUN COUNT: 'The Coin Flip Judge?' (arXiv 2606.13685, 2026-04-23) measured repeated identical LLM evals flipping pairwise preference 13.6% of the time (28% of questions >20% flip); 11 repeated trials are needed for a majority vote to recover the 50-trial reference verdict at 95% probability, 15 for high-variance questions. >=5 runs is a SMOKE READ, not a verdict — report >=11. INDEPENDENCE: COLM 2026 (arXiv 2604.07650) shows 18 LLMs / 6 families induce 'latent entanglement' from shared pretraining-distillation-alignment — correlated reasoning and SYNCHRONIZED FAILURES — so three vendor families are NOT three independent judges; disentangle by decode style (system-prompt / no-system-prompt / different temperature) and keep one human read of the raw panel. SELF-PREFERENCE: 'Play Favorites' (arXiv 2508.06709) — judges rate their own and same-family outputs higher, so a JUDGE must never share a family with the GENERATOR of a candidate. SCORING: 'Reliability without Validity' (arXiv 2606.19544) measured exact-match agreement OVERSTATING discriminative ability by 33-41pp of kappa; and protocol choices alone move reported accuracy 0.551 -> 0.899 without changing a single verdict (arXiv 2606.00093). So: report a chance-corrected score (Cohen's kappa / ICC) AND the full protocol (runs, families, styles, temperature, abstentions) so the number is reconstructible. MODE% ALONE IS NOT A VALIDITY STATISTIC — an unreconstructible score is not a score. Third-person framing STAYS, now empirically justified (GEM 2026, aclanthology 2026.gem-main.45: first-person framing biases judgement toward the narrator). THE PROBE KILLS CHEAPLY; IT DOES NOT ANOINT — it remains an operationalisation of GATE 0h, not a substitute for the invoice test or the founder's ear. (2) AVAILABILITY TRUTH for .ai/.io: Identity Digital serves BOTH and its port-43 whois is being SUNSET (the whois.nic.ai banner redirects to RDAP) — this CONFIRMS TOOLS.truth's hardcoded rdap.identitydigital.services. NEVER port-43 whois for these TLDs (DO_NOT_REVERT #5's catastrophe on non-.com, now structurally guaranteed). Registry-grade replacement for the false-positive-prone MCP/DNS filter: Porkbun checkDomain with dryRun:true (porkbun.com/llms-full.txt) — availability + price in one idempotent call. (3) LEGAL GATE additions: WIPO D2025-3938 NARROWED the 'brandable portfolio' defence — a bought brandable must convey a linguistic or conceptual meaning INDEPENDENT of the complainant's mark, or it reads as cybersquatting on that mark. This is precisely our AFTERMARKET buy path, so it is now a named risk. UDRP also gained an EXPEDITED track from 2026-03-09 (~1 month to decision; $4,000 complainant / $2,500 respondent) on a record 6,200+ case year. Budget the fast lane. DAI2025-0060: for an AI-native brand the .ai string itself can become part of the protected mark. (4) MARKET: Verisign .com wholesale +7% to $10.97 effective 2026-11-01 — the defensive-variant math ('grab a typo or two, ~$10 each') moves with it; .web was delegated 2026-07-22 and is coming to market — review the novelty-TLD blacklist stance when it launches; the ICANN 2026 round closed 2026-08-12 with 1,600+ strings, Reveal Day ~Oct 2026 — re-sweep then for strings that collide with live client names (WIPO is the exclusive LRO/SCO objection provider for this round). USPTO first-action output hit a record with inventory declining (bulletin 2026-04-23) — the register is saturated but the pace has turned in our favour; still clear early. TOOLING: add NameBuddy.ai (registry-checked, no affiliate 'premium' pricing — the MCP-false-positive failure is its whole pitch), Spamzilla's free tier for small expired-name vetting batches, and the DomainsGPT pattern worth copying in our own domain screen — show a third-party valuation BESIDE the live registration price. Dan.com is GONE (listings migrated to Afternic); the aftermarket path is GoDaddy/Afternic only.

Pipeline v2 after the zeitgeist sweep (full brief: naming/context/sota-methodology-2026.md). LAW: LLMs are PROFILERS not generators - generation comes from morphemes + phonesthemes with a REQUIRED one-line morpheme story per candidate; zero-anchor coinage = linguistic averaging (the Axera/ondel mush lane), DEMOTED to last resort. New gates next to 0h: (1) say-it-twice-fast phonetic gate (stress/spelling ambiguity kills - ONDIN died here), (2) 3-second category signal, (3) 15-second story test ('it sounded cool' = fail), (4) zeitgeist check vs the 2023-26 era winners (real English words with meaning, serious composites, iconic short oddballs, first names, animals - per Crunchbase era map 2025-05; anti-patterns -ly/-ify, misspellings, zero-anchor -a/-ia mush). BRIEF first: taste anchors from the founder (brands loved/hated, names he'd buy, his own instinct word - his 'onair' = the always-live broadcast register) gate every generation cycle; quantity >=50/cycle, 2 cycles, names presented as tools not verdicts (Lexicon). Sources: Tanj 2026-03 (linguistic averaging), Focus Lab x Placek 2026-01 (Identify-Invent-Implement, AI-as-profiler, 107 linguists + 7 in-house LMs, URLs are area codes), NameStormers 2026-01 (name slop filter), River+Wolf 2024/26 (prompt tonality effect, lateral paths, VC-portfolio zeitgeist mining), Klink JCR + JCPY 2024 (phonestheme evidence).