Why V2
V1 worked. It generated pronounceable names, screened trademarks, and checked domains over RDAP. It also had the wrong objective function — and no amount of tuning fixes an objective that optimizes for the wrong thing.
Why V1 wasn't enough
The first engine treated naming as a supply problem: generate many candidates, filter out the broken ones, sort by a composite score, and let domain availability drive the ranking. That produces usable names. It does not produce names like the ones founders actually admire — the short, confident, one-idea names that established companies carry. V1's four styles (invented, real words, compound, descriptive) described mechanisms, not aesthetics, and its lexicon was one undifferentiated pool: asking for Nordic minimalism or Latin gravitas wasn't expressible at all.
The deeper problem was honesty. A single opaque number ranked every candidate, domain availability could outweigh whether a human could pronounce the name, and nothing in the output explained why a name ranked where it did. A tool that scores names owes its user the reasoning.
The tempting fix was more heuristics — sharper filters, better penalties, a longer blocklist. None of it would have mattered, because you cannot tune your way out of a wrong objective: scoring availability more cleverly still scores availability. V2 starts one level up. It redefines what good means — a name a founder would believe a company could be built on — and only then rebuilds the machinery that pursues it.
Design goals for V2
- The benchmark is brand quality, not quantity — the test is whether a name could sit beside the companies founders admire, not whether a .com is free.
- Aesthetic direction must be a first-class input: a founder chooses a naming philosophy, and every layer of the engine obeys it.
- Every score must be defensible: deterministic, inspectable, and phrased no more confidently than the method deserves.
- The engine must explain its own reasoning in human terms, derived from the same arithmetic that produced the rating.
- Claims about linguistic origin must be true. An invented word is not Old Norse because it borrowed a root.
Architecture
The rebuild became much simpler once one decision was made: separate generation from judgment. Generation proposes; scoring judges. Once those responsibilities stopped sharing code paths, everything else — the philosophies, the gates, the explanations — followed naturally. The diagram is the result; the separation is the idea.
Why deterministic scoring instead of AI-only generation
An LLM can propose names; V2 still supports that, strictly bring-your-own-key. But delegating judgment to a model would make the product's core claim unauditable. When quality is a number a founder might act on, the number needs to come from arithmetic someone can read: the same input always scores the same, a change in the rules is a reviewable diff, and every displayed claim traces to a threshold that either was or wasn't crossed.
So the division of labor is strict: models may only ever propose. Every candidate — human, local, or model-proposed — passes the same deterministic gates and is scored by the same deterministic rubric. The judge never generates and the generator never judges.
How we rebuilt it
The rebuild ran as four moves, in order. Each one was only possible because of the one before it.
- Separate generation from judgment. Six generation strategies and an optional model on one side; a single deterministic rubric on the other. Every candidate faces the same judge regardless of where it came from.
- Replace mechanical styles with philosophies. Styles described how names were assembled; philosophies describe what names should feel like, and bias generation, pools, scoring, and prompts to match.
- Rewrite scoring around deterministic heuristics. Eleven inspectable dimensions, a composite with disclosed structure, and qualitative bands where false precision used to be.
- Make every score explain itself. Once judgment was arithmetic, explanations came almost for free — each claim traces to a threshold the name actually crossed.
The scoring system
Every mechanism in this section exists because something simpler failed first. Rather than describe the machinery, this section keeps each failure attached to its fix.
›Naming philosophies & weighted generation
- Warum sie existiert
- The first tell was a hack I had written myself: when a brief sounded premium, V1 quietly boosted the materials word-list. It looked reasonable until I noticed what it implied — aesthetics were already steering generation, just dishonestly, buried in keyword matching. If taste was going to steer the pool, it deserved to be a first-class input instead of a side effect. That is what a philosophy is.
- Architektur
- So taste became declarative. A philosophy states, in one config, everything the old hack smuggled through keywords: which generation strategies run and in what proportion, which root banks feed the pools and how heavily, how scoring bends for the mode, what an optional bring-your-own-key model is told. The surprise was what this exposed about the lexicon. I had thought of the banks — Norse, Celtic, Latin, Greek, nature, science, architecture, geography, astronomy, mythology, two English registers — as an asset the whole engine shared. Flattened into one pool, they were actually a ceiling: every request got the same soup with different sorting. Nordic minimalism wasn't unsupported in V1. It was inexpressible.
- Einschränkungen
- Runs with no philosophy selected reproduce V1 behavior exactly; the wire format is purely additive. The first-run UI shows six primary directions and folds the rest behind a disclosure, because seventeen equal chips is a wall, not a choice.
- Trade-offs
- Weighted pools mean deliberate repetition inside the candidate soup — the dedupe happens downstream by design. And modes share strong stems, so two adjacent philosophies can surface the same excellent name; the engine prefers honest overlap to artificial variety.
›Linguistic provenance — why honesty matters
- Warum sie existiert
- The origin column began as decoration. Generation draws from named root banks, so labelling each result with its bank felt free — until I read a row the way a stranger would. The engine had grown a coinage off a Norse stem and was presenting the entire invented word as Norse, a claim no linguist would sign. The bank audit that followed made it worse: one word filed under Greek turned out to be Māori. The engine was making claims it could not defend, and I had written them.
- Architektur
- The wrong assumption was treating origin labels as metadata. They are claims — a stranger reads 'Norse' as a fact about the word, not a note about which file the stem came from. So the labels now match the evidence, in three tiers: a word that is a curated bank entry shows its tradition plainly (Vidar · Norse); a coinage built on a bank root says inspired (Norse-inspired), never native; a pure invention claims nothing, and no label renders at all. The audit cut the other way too — bank entries that could not defend their own filing were removed, not relabelled.
- Randfälle
- English is the awkward tier: a coinage grown from an English stem is just a coined word, so the inspired label is suppressed there — 'English-inspired' would be noise pretending to be information.
- Trade-offs
- Three honest tiers convey less romance than one dishonest label. The record accepts that.
›Modern Brand Fit — and its limitations
- Warum sie existiert
- Iconic technology brands cluster into a measurable statistical shape — length, syllable count, vowel balance, how the name opens and how it ends — and a candidate can be scored by its distance to that profile. The dimension was born under the name DNA fit, and that name was itself a bug: DNA implies destiny, and the measurement is resemblance. It shipped renamed to what it actually is.
- Architektur
- Not trusting my own taste became the other half of the design. The profile is derived at load from a fixed corpus of well-known company names — measured, never hand-tuned — so if my sense of the modern shape is wrong, the corpus outvotes me, and editing the corpus updates the profile with no code change. What renders anywhere a person reads it is Modern Brand Fit: resemblance to the current shape of successful naming, nothing more.
- Einschränkungen
- Its weight in the composite is deliberately modest. Fitting the profile is evidence of familiarity, not a cause of success — plenty of great companies violate the shape, and the shape itself will drift as fashions do.
- Trade-offs
- This dimension encodes survivorship: it studies winners. That is exactly why it is one voice among eleven rather than the headline, and why the corpus itself stays out of both the UI and this article.
›Brandability bands versus numeric scores
- Warum sie existiert
- V2's first scoring pass shipped a number with one decimal place. I stopped trusting it the day the evaluation set came back: a name from the engine's own list of negative examples was carrying an 89, and nothing about the number 89 invited the question. That is the real cost of false precision — it does not just overstate confidence to readers, it stopped me from interrogating my own scorer.
- Architektur
- Bands are the admission, not a styling choice. Brandability now presents as Exceptional, Strong, Promising, or Weak — as coarse as the method actually is — while the precise number retreats to the expanded breakdown, labelled a heuristic, next to every component that produced it. The evaluation forced one more piece I had not planned: a floor rule, because I watched a name with hopeless credibility climb the composite on phonetic polish alone. A name that fails believability can no longer buy its way into a high band.
- Einschränkungen
- The composite is weighted so the cognitive dimensions — credibility, pronunciation, memorability, spelling — each individually outweigh the supporting signals like visual balance and domain availability. Domain availability influencing brand judgment was a V1 mistake this rebuild reversed.
- Trade-offs
- Bands lose resolution at the boundaries; two names a point apart can wear different labels. Accepted — boundary noise on an honest scale beats false confidence on a precise one.
›International pronunciation testing
- Warum sie existiert
- V1 tested pronunciation against the founder's one selected market. That looked reasonable until I scored names intended for a global brand and realized the engine had only ever asked a single phonology for permission. A name that travels has to survive sound systems it was never designed for.
- Architektur
- My first sketch was a veto: any market that chokes kills the name. Thinking that through against the strictest profile ended it — a phonology that forbids consonant clusters and closed endings would execute half of the best English brand shapes at the gate. So each profile encodes what its speakers tolerate — cluster limits, awkward letter patterns, ending expectations, letters that carry a pronunciation tax — and the dimension runs every candidate against all of them and averages how cleanly it survives. A high score means nobody choked.
- Randfälle
- The profiles are phonological heuristics, not semantic checks. They will not catch a name that is perfectly pronounceable in another language but means something unfortunate there. That check remains human.
- Trade-offs
- Averaging is forgiving by design; a name brutal in exactly one market can still score moderately. That is the standing cost of refusing to let one profile govern the pool.
›Competitive benchmarking & the phone test
- Warum sie existiert
- Not everything from V1 failed. Two mechanisms survived the rebuild with their roles enlarged: screening candidates against known marks and brand stems, and the phone test — if someone hears the name once over a phone call, will they spell it right? The rebuild's job was to stop burying their verdicts under a single opaque number.
- Architektur
- The phone test simulates the failure modes of hearing: ambiguous digraphs, silent endings, plausible alternate spellings. It reports pass or fail with the alternates it fears. Conflict and trademark screening stay deterministic — string similarity, phonetic keys, and class relevance against a bundled dataset — and stay clearly labelled as informational screening, not legal clearance.
- Trade-offs
- The trademark dataset is a screening aid; a real clearance search is a lawyer's job and the UI says so in as many words.
Phonetic families
Phonetic families exist because of a rule I was proud of. The engine hard-rejects mechanical root-plus-suffix coinage — its best defense against lazy names — and that same rule made the most useful gesture impossible. When a founder almost loves a name, the right answer is its siblings: Arvo wants Arven, Arvon, Arvia, Arvik on the table. Siblings are root-plus-suffix by definition, so the quality bar was blocking exactly the feature users needed, and V1's only answer was regenerate and hope.
The generator extracts the stem, then builds variants from suffix registers — coherent ending systems, one register per family. The temptation was to mix registers for variety, but siblings share more than a prefix; a family that borrows endings from everywhere stops sounding like a family. So each one commits to a single system and lets depth, not variety, make the case.
The first draft of the family generator made its own mistake: it filled a quota — six variants per family, padded when the stem could not honestly support six. Testing killed that too; on many stems, most of the padding failed the same phone test the main table enforces. Now every variant independently passes five gates — the phone test, credibility, pronunciation, spelling clarity, and collision screening — and there is no quota to fill. Three strong variants beat five padded ones, and an empty family is a real answer: it means the seed is already the best member of its family.
The one deliberate exemption
The main engine hard-rejects mechanical root-plus-suffix constructions — that rule is its defense against lazy coinage. A family generator exists to build root-plus-suffix systems on purpose, so it exempts that single rule and keeps every other quality signal intact. The exemption is scoped to families and documented in the code; nothing else in the engine can use it.
Explainability
Explainability was not on the original V2 list. It forced its way on when the composite score stopped being trustworthy — first through the evaluation failures, then through a quieter observation: when I could not say in one sentence why a name scored 91, I had no way to notice when 91 was wrong. A score you cannot narrate is a score you cannot audit.
So every explanation is assembled, not written. A claim like 'easy to pronounce after hearing it once' appears only when the pronunciation score crossed its threshold, and claims are ordered by their actual weighted contribution to the brandability composite — the engine explains the same arithmetic it scored with, which means the explanation cannot drift from the rating. The first draft then taught one more lesson: equally clean names earned identical bullets, and five identical bullets across a table read as template text even though every word was derived. The fix reserves the final slots for the facts specific to that name — its linguistic root, its phone-test result, live domain availability.
The same honesty rules bind the prose: no bullet implies business outcomes, origin claims respect the three provenance tiers, and a weak name gets no manufactured case — it points to its score breakdown instead of reaching for compliments.
- Reads as an established company, not a coinage.
- Short, concrete, and easy to recall.
- Easy to pronounce after hearing it once.
- Built on an authentic Latin root.
- Passes the phone test — spellable after one hearing.
Evaluation
A scoring engine that has never been tested against human judgment is an opinion generator. V2 ships with a fixed evaluation set — thirty successful modern brands (half deliberately outside the benchmark corpus), thirty intentionally weak or awkward names, and twenty difficult edge cases — and the first full run made the case better than any argument could: a name sitting in the engine's own negative examples, the class its prompts explicitly ban, came back scored 89 out of 100. The scorer and the manifesto disagreed, and the evaluation was the instrument that caught it.
What the numbers say
- The successful and weak cohorts now separate by roughly twenty-seven points of median brandability, and no weak name reaches the Exceptional band.
- The 89 was repaired with principled rules, not corpus tuning: names with no vowels, no consonants, or AI-buzzword tails can no longer be rescued by their other components, and the soft-coinage class the manifesto bans is scored the way the manifesto talks about it.
Where it still fails
- Sci-fi letter sludge — invented names with plausible letter statistics but no linguistic soul — can reach the Strong floor. Letter-pattern heuristics genuinely cannot distinguish this class from legitimate coinages; fixing it needs phonotactic modeling or a judging model, and the record prefers the disclosed gap to the opaque fix.
- Eight genuinely successful brands score Weak on this rubric — the consonant-wall, added-suffix, and long-compound classes. That is a deliberate design stance, not an accident: the engine optimizes for one specific shape of name, and those companies won carrying names this rubric would have advised against.
- The evaluation set is fixed and public inside the repo, which means future tuning could overfit to it. The guard is procedural: scorer changes are argued from rules, and the set exists to catch regressions, not to be climbed.
Trade-offs & unsolved
V2 is a set of positions, and positions have costs. The ones that matter are recorded here, next to what V1 did, so the change is auditable.
What changed from V1 to V2
| Dimension | V1 | V2 |
|---|---|---|
| Objective | Available, pronounceable names | Names that could carry a company; availability is one signal |
| Direction control | 4 mechanical styles | 17 philosophies biasing generation, pools, scoring, and prompts |
| Lexicon | One flattened pool | 12 root banks, weighted per philosophy |
| Headline score | One opaque number | Qualitative band; numbers in the breakdown, labelled heuristic |
| Provenance | None | Authentic root / inspired / no claim — never a false language label |
| Alternatives | Regenerate and hope | Phonetic families behind five independent quality gates |
| Reasoning | None | Score-derived explanations on every candidate |
| Validation | None | Fixed 80-name evaluation set, failures published |
What remains intentionally unsolved
- Semantic meaning across languages. The engine tests whether a name can be pronounced abroad, not what it means there. A native-speaker check is irreplaceable and pretending otherwise would be the exact false confidence V2 exists to remove.
- Predicting success. Nothing in this engine forecasts whether a company will work. The Future Unicorn direction is a shape heuristic and is labelled as one; the moment a naming tool implies it can pick winners, it has left engineering for astrology.
- Taste. The rubric encodes one defensible aesthetic — the short, credible, one-idea name. Founders whose taste runs elsewhere are not wrong, and the tool does not pretend its rubric is the only valid one.
- Legal clearance. Screening reduces surprises; it does not replace counsel, and the interface says so wherever a trademark result appears.