A Three-Layer Framework for Measuring Names and Its Census Application on a Token Launchpad

• The study counts all 513,647 naming attempts, including tokens that never traded.Source: Highlights · p. 2
• A three-layer framework extends fluency measurement to Chinese names.Source: Highlights · p. 2
• A triple-benchmark protocol evaluates the reliability of LLM-assisted content coding.Source: Highlights · p. 2

Introduction

• Existing alphabetic-script fluency measures do not transfer to Chinese names, leaving Chinese-named assets and corpora largely outside current measurement research.The passage notes that Chinese names require measures suited to their writing system rather than letter-sequence pronounceability or bigram frequency.Source: §1 · p. 4
• Manual cultural coding is designed for small samples, while large-scale LLM coding still lacks established reliability practices for repeated coding and adjudication.Existing studies typically cover hundreds of names, far below issuance populations of hundreds of thousands.Source: §1 · p. 4
• Treating each name in isolation misses its relations to prior reuse and simultaneous semantically similar competitors, dimensions not widely applied in asset-naming research.The passage identifies name-population position as part of a name’s economic meaning.Source: §1 · p. 4
• The paper addresses all three measurement gaps in the Four token launchpad on BNB Chain, where names are primary token identifiers in trading interfaces and listings.The study restricts its object to textual attributes, with the name at the core, and leaves image attributes to future research.Source: §1 · p. 5
• The launchpad makes naming a nearly costless, unrestricted text decision fixed before trading outcomes, supporting study of names apart from fundamentals.The introduction also describes naming as continuous and high-frequency, with Chinese names forming a substantial share of the corpus.Source: §1 · p. 5

1. Introduction

• Naming is observed without a survivorship filter, and unconstrained reuse across times and addresses enables direct measurement of inter-name relations.In contrast, rejected naming attempts are unobservable in equity or fund research, while uniqueness rules make such reuse observations impossible for stock tickers and domain names.Source: §1 · pp. 5–6
• Tokens share contract code, supply schedule, and pricing mechanism, leaving names among their few substantive differences without confounding from other attributes.The passage contrasts this homogeneous setting with differences in governance structure and underwriting quality.Source: §1 · p. 6
• The framework measures names along three mutually orthogonal dimensions: processing fluency, cultural reference, and inter-name relations.It extends fluency measurement to Chinese-named assets using indicators grounded in Chinese characters.Source: §1 · p. 6
• Large-language-model coding with strict reliability testing scales cultural content analysis to census size.The framework also extends processing-fluency measurement to Chinese-named assets for the first time.Source: §1 · p. 6
• The study characterizes statistical regularities in name reuse and name-space expansion using a fully observed naming population.The framework, reliability protocol, and population benchmarks are presented as transferable to other markets and corpora.Source: §1 · p. 6
• The framework, reliability protocol, and population benchmarks are transferable to other markets and corpora, supporting future causal tests of names and market outcomes.Source: §1 · p. 6

2.1 Institutional setting

• The setting is the four.meme token launchpad on BNB Chain.Source: §2.1 · p. 6
• Token issuance deploys a fixed supply of one billion units onto a bonding curve, with liquidity migrating at 17.64 BNB.The threshold is constant throughout the research window.Source: §2.1 · p. 6
• Issuance requires no review, queuing, or disclosure, and names are not constrained by uniqueness.Source: §2.1 · p. 6
• Repeated names are observable across times and circumstances, making name content separable from naming circumstance.There were 237,937 unique names across 513,647 attempts; 71,585 names were reused, with a maximum reuse count of 1,130.Source: §2.1 · pp. 6–7
• The paper calls the set of tokens sharing a normalized name a name lode.Source: §2.1 · p. 7

2.2 Cohort and data sources

• The fixed cohort includes 513,647 tokens created from February 1 through June 30, 2026.The cohort was determined before data collection and enumerated from factory creation events.Source: §2.2 · p. 7
• 58,816 tokens (11.4%) produced no trades and were retained to avoid survivorship bias in statements about the naming population.Source: §2.2 · p. 7
• A two-source pipeline combines on-chain event discovery with attribute and trade data, and cross-checks token enumeration across sources.The trade records total 52,066,542.Source: §2.2 · p. 7
• Names come from on-chain creation events that cannot be tampered with or edited after the fact.The paper identifies this as the data-level reliability foundation for the object studied.Source: §2.2 · p. 7

2.3 Census completeness audit

• 90.2% of tokens missing market snapshots had zero trades; their migration rate was 0% and mean trade count was 0.11.The missingness therefore concentrates on tokens with no market activity and is described as structural non-measurement rather than random data loss.Source: §2.3 · p. 8
• Market-field gaps do not affect naming-side completeness: the name field remains fully measurable, including for zero-activity tokens.Source: §2.3 · p. 8
• Creator address, name, migration flag, internal-market aggregates, and trade counts are covered at 100%; holder lists at 97.09% and market snapshots at 87.31%.Source: §2.3 · p. 7, Table 1 · p. 8
• The corpus contains 511,664 non-empty names, representing 237,937 unique names; 41.7% are Chinese, 44.3% English, and 11.8% mixed Chinese–English.Another 2.3% are categorized as other; 1,983 empty names are listed separately.Source: §2.3 · p. 8
• 35,677 creator addresses issued tokens, and prolific addresses contributed 95.3% of all tokens; the top address issued 9,750 tokens in five months.Source: §2.3 · p. 9

2.4.1 The form layer: fluency measures adapted to the Chinese writing system

• The form layer uses 38 fields: 20 for Chinese, 12 parallel English fields, and six cross-language fields, because alphabetic-script fluency measures do not transfer directly to Chinese names.Chinese fields cover phonological, graphemic, and lexical facets; English fields follow the principle of the same structure but different sources.Source: §2.4.1 · p. 9
• Only 6.65% of Chinese names differed in tone entropy between contextual and character-by-character readings, with a mean difference of 0.117, setting the measured error bound for this dimension.The paper presents the form layer as the framework’s measurement-error benchmark.Source: §2.4.1 · p. 10
• The six core Chinese form dimensions have mean absolute pairwise Spearman correlation 0.098 and first-component variance 25.2%, while their English counterparts are 0.277 and 37.7%.For pure Chinese names, syllable count and character count are numerically identical and should not enter a model simultaneously.Source: §2.4.1 · p. 10

2.4.2 The reference layer: codebook, double-blind coding, and LLM expansion

• The reference layer combines rule-based categories with six judgment-coded dimensions, establishes a double-blind human gold standard on 408 names, and expands coding to all 237,937 unique names using an LLM.The codebook supplies decision rules, examples, and boundary cases; prompts were calibrated separately from validation before full-scale expansion.Source: §2.4.2 · pp. 12–13
• After codebook alignment, second-round Krippendorff’s α ranged from 0.756 to 0.982 across six dimensions, with D1b at 1.000; both rounds are reported to avoid overstating independent agreement.First-round agreement was low for D2 (0.248) and D5 (0.313).Source: §2.4.2 · p. 12
• D1, D3, and D5 were confirmatory, while D2, D4, and D6 were exploratory because of machine self-consistency or human confidence-interval shortfalls.D2's machine self-consistency was 0.655; D4 and D6 failed the human confidence-interval threshold.Source: §2.4.2 · p. 13
• On D4, coders disagreed on 14 names, with 12 disagreements in the same direction, including over festival modifiers without a specific event reference; D4 was exploratory.The codebook did not cover these boundary cases, and the procedure allowed no third coding round.Source: §2.4.2 · p. 13

2.4.3 The relation layer: order, crowding, and variation

• The relation layer captures lode order and reuse, semantic-neighborhood crowding, lexical variation, and launch density; 6.4% of names had precursor anchors, and retrieval recall was 97.72%.Crowding reflects recent names sharing content words; the retrieval approximation's missed pairs were near the similarity threshold, biasing imitation coverage conservatively.Source: §2.4.3 · pp. 14–15
• Of 220,045 distinct-name lodes, 148,460 had one token and 71,585 had multiple; roughly one in three names was reused, with a maximum count of 1,130.Tokens sharing a normalized name formed a lode; multi-token lodes indicate reuse over time, often by different creators.Source: §2.4.3 · p. 14

2.4.4 Aggregation and reproducibility

• Layer subindices use covariance-inverse-weighted GLS with feature directions set programmatically; code, codebook, annotations, and statistic scripts are released, while raw trade-level data require controlled access.The measurement takes the name, rather than the symbol, as its object because symbols are often abbreviations or truncations and overlap substantially with names.Source: §2.4.4 · p. 15

3.1 Structure of the measurement framework

• Pairwise Spearman correlations among the three layers were no greater than 0.11, supporting their nonredundant measurement of name attributes.The reported correlations were 0.11 for form–reference, −0.03 for form–relation, and 0.00 for reference–relation.Source: §3.1 · p. 16, Figure 2 · p. 16
• Authority borrowing (D1) appeared in 45.60% of mixed Chinese–English names, versus 23.30% of Chinese and 19.63% of English names; seasonal events and financialized blessings were also more common in Chinese than English names.The respective Chinese–English positive rates were 2.94% versus 0.84% for seasonal events and 3.86% versus 0.85% for financialized blessings.Source: §3.1 · p. 16, Table 4 · p. 17
• Referential co-occurrence tracked semantic relatedness: ironic self-reference with cultural memes had a 4.21 lift, while cultural memes with financialized blessings had a 0.46 lift.Most names carried zero or one reference dimension (63.77% and 31.58%, respectively); two dimensions occurred in 4.52% and three or more in 0.13%.Source: §3.1 · p. 17

3.2 The expansion law of the name space

• Cumulative unique names followed V = 1.24 · N^0.921 (R² = 0.999), indicating diminishing marginal novelty as naming attempts accumulate.The exponent near 0.92 also indicates that the name space remains far from saturated.Source: §3.2 · p. 18, Figure 3 · p. 18
• The share of new names fell monotonically from 52.0% in February to 37.8% in June, directly reflecting the subunitary Heaps’ exponent.The study describes the decline as roughly three percentage points per month in a market where naming is costless and uniqueness is unconstrained.Source: §3.2 · p. 18, Figure 3 · p. 18
• The Chinese share fell from 44.5% in February to 36.8% in March, then rose to 50.3% in June, while the English share fell from 47.3% to 38.2%.Because the language groups occupy periods with different market conditions, the authors caution that cross-language comparisons should control for time; the observed drift relies on continuous observation of issuance.Source: §3.2 · pp. 18–19

3.3 The distribution of name reuse

• 67.5% of 220,045 lodes contained one token, while reuse among the remainder was heavy-tailed, with Hill tail exponents of 2.19–2.67.The reported exponent range implies a finite mean but divergent variance in lode size.Source: §3.3 · p. 19
• Intervals between successive uses of the same name were bimodal, with a median under three minutes, and naming activity was reported as sensitive to external events.Around the Spring Festival Gala broadcast, the study examined hourly creation counts and the share of festival-related names.Source: §3.3 · p. 20
• The largest name lode contained 1,130 tokens, while 67.5% of names were used exactly once and the median interval between successive same-name uses was under three minutes.The reuse distribution is presented as a survival function of lode size on log–log axes.Source: Figure 4 · p. 20, §3.3 · p. 21

3.4 The two-scale temporal structure of copying

• Same-creator copies had a 0.7-minute median interval, versus 5.6 minutes for cross-creator copies (p < 0.001), separating batch deployment from imitative adoption.The interval difference was highly significant under a Mann–Whitney test; same-creator repetition reflects rapid redeployment by one operator, while cross-creator repetition reflects others adopting names.Source: §3.4 · p. 21
• Among 71,585 reused lodes, those with more than five creators were 16.8% of lodes but contained 50.7% of tokens and spanned a median 666.5 hours.Single-creator lodes were 9.2% of lodes and contained 4.7% of tokens; lodes with 2 to 5 creators were 74.0% of lodes and contained 44.6% of tokens. The pattern indicates that most reuse is cross-actor adoption extending over weeks, rather than instantaneous batch deployment.Source: §3.4 · p. 21

3.5 Naming responses at event time

• Within six hours of the Gala broadcast, issuance reached 4,128 tokens, roughly 2.8 times the preceding six-hour count, while festival-related names made up 12.3%.The festival-name share was 17% to 18% during the twelve hours before onset and fell to 1.7% in the following window, so the volume surge did not indicate a shift toward festival vocabulary.Source: §3.5 · p. 22

4.1 Portability of the framework

• Form- and reference-layer measurement uses the name string and a codebook, making it portable across naming corpora; relation measures require name reuse and creation timestamps.Where uniqueness is enforced, lodes and naming order do not exist, but crowding and lexical variation remain constructible.Source: §4.1 · p. 23

4.2 Research potential of the data

• The Heaps exponent, heavy-tail exponents, and new-name-share decay offer a testing ground for models of lexical innovation and cultural transmission.Unlike natural-language corpora, this dataset traces each word’s birth and adoption item by item.Source: §4.2 · p. 24
• The census traces copying at two scales—0.7 minutes within creators versus 5.6 minutes across creators—and records adoption actors and timing item by item.Its second-level timestamps support studying imitation, diffusion, and weeks-long adoption processes.Source: §4.2 · p. 24
• The new-name share declines by roughly three percentage points per month, while the largest 1% of lodes contain 17.4% of tokens.These population patterns characterize a name space without property-rights constraints and provide input to questions about uniqueness or property-rights rules.Source: §4.2 · p. 24

4.3 Limitations

• Validity evidence covers convergent, discriminant, and structural validity, but not predictive validity, which awaits a companion study.The six coding dimensions are tailored to this setting and may require a rebuilt codebook elsewhere; the reliability protocol, rather than dimension content, is what transfers.Source: §4.3 · p. 24
• The descriptive regularities come from one platform and one time window, so their stability across platforms and periods awaits parallel censuses.Reproducibility of LLM coding is also constrained by model availability, which remains beyond the authors’ control despite archived prompts and versions.Source: §4.3 · p. 25

5. Conclusion

• The study builds a census-level dataset of 513,647 naming attempts and proposes a three-layer framework for measuring names.The data cover five months on the four.meme token launchpad on BNB Chain.Source: §5 · p. 25
• The form layer uses 38 Chinese-adapted deterministic fields, the reference layer combines human coding with LLM expansion, and the relation layer measures names’ positions in the population.Reference-layer reliability is reported across human double-blind agreement, human–machine replication, and machine self-consistency.Source: §5 · p. 25
• The three measurement layers are nearly orthogonal in the data and each carries independent information.Source: §5 · p. 25
• The census finds Heaps’ law growth, declining new-name adoption, language-composition drift, heavy-tailed reuse, and repeated uses spanning distinct creator-level and cross-creator timescales.The reported patterns include V = 1.24·N^0.921 (R² = 0.999), a decline in new-name share from 52.0% to 37.8%, and reuse tail exponents of 2.19 to 2.67.Source: §5 · p. 25
• Repeated use of the same name separates into 0.7-minute within-creator batch deployment and cross-creator adoption starting at 5.6 minutes and extending over weeks.The median interval between successive uses is under three minutes.Source: §5 · p. 25
• Cultural events trigger immediate quantity responses, while festival-vocabulary use follows the calendar, making the responses temporally separable.Source: §5 · p. 25
• The framework, Chinese fluency yardstick, and LLM reliability protocol provide reusable infrastructure, while the actor- and time-identifiable census offers an object of study for lexical innovation, imitation, diffusion, and name-property-rights design.Source: §5 · pp. 25–26

CRediT authorship contribution statement

• Dingding Cao led conceptualization, methodology, software, data curation, formal analysis, and the original draft; the other authors’ roles are also specified.Source: CRediT authorship contribution statement · p. 28

Funding

• The work reports support from the Baoshan University Doctoral Research Startup Fund Program 2025 and a Yunnan Province science and technology plan youth project.Source: Funding · p. 29

Declaration of competing interest

• The authors declare no competing interests and no positions in any token in the study.Source: Declaration of competing interest · p. 29

Data and code availability

• Replication materials are archived on Zenodo, while raw trade-level data are available under controlled access.The archived materials include the census master table, annotations and codebook, implementation code, and character reference table; the authors state that every reported number can be independently recomputed.Source: Data and code availability · p. 29

Declaration on the use of large language models

• The LLM was used for full-scale expansion of the reference-layer coding and for language polishing and code translation or refinement; the authors wrote the paper text.Source: Declaration on the use of large… · p. 29

Ethics statement

• The research uses public on-chain data and public APIs.Source: Ethics statement · p. 29