A cross-domain tropical species dataset with Chinese vernacular names and CITES source links

# Abstract
Source: [https://arxiv.org/html/2606.03156](https://arxiv.org/html/2606.03156)
A cross\-domain tropical species dataset with Chinese vernacular names and CITES source links Jeff Wang NEXLY LLC, United States

Correspondence:[jeff@tropicals\.cn](https://arxiv.org/html/2606.03156v1/mailto:[email protected]) ORCID:[0009\-0001\-2905\-8439](https://orcid.org/0009-0001-2905-8439)

We describe a versioned cross\-domain dataset of 410,499 active tropical species \(working snapshot 2026\-04\-20\) spanning three applied subdomains —tropical\_plants,tropical\_aquatic, andtropical\_pets— that share a commercial and regulatory life cycle but are distributed across kingdom\-organised biodiversity infrastructures\. The resource joins taxonomic identifiers from GBIF, Plants of the World Online, iNaturalist, NCBI Taxonomy, the Catalogue of Life and the Encyclopedia of Life, and adds three original layers: a cross\-domain ontology that re\-segments taxa along trade and husbandry contexts; a Chinese vernacular layer with explicit per\-name provenance under a typology that excludes unverified machine\-generated proposals; and a CITES source\-linkage layer connecting each taxon to its Species\+ entry\. Chinese vernacular*coverage*— the proportion of taxa carrying a CJK Chinese name distinct from the scientific binomial — reaches 99\.50 percent \(408,456 of 410,499; full\-population count\)\. Coverage characterises completeness, not name\-translation accuracy; the latter is bounded by the four\-level provenance typology and is the subject of a preliminary internal review reported here, with a blind external audit identified as the principal open item\. Upstream content is referenced by stable identifier only for the original\-contribution layers, supporting CC\-BY 4\.0 reuse\. The dataset is deposited on Zenodo \(10\.5281/zenodo\.20377811\)\. This preprint is the canonical v1\.0 description of the dataset’s current state; future Data Descriptor submission is anticipated but is contingent on the validation and release\-engineering items listed in §Limitations\.

## Keywords

biodiversity informatics; tropical species; chinese vernacular names; cites source linkage; darwin core; cross\-domain compilation

## Background & Summary

The systematic documentation of species rests on a binomial nomenclatural system now realised through a small number of large public infrastructures\. The Global Biodiversity Information Facility mobilises occurrence records and a taxonomic backbone as a global open infrastructure for biodiversity data \[1\]\. Plants of the World Online provides curated botanical nomenclature, distribution, uses and conservation status \[2\]\. iNaturalist contributes a community\-observation graph with taxonomic identifiers \[3\]\. NCBI Taxonomy maintains a curated classification linked to the public sequence databases \[4\]\. The Catalogue of Life integrates source checklists into a single global species list \[5, 18\]\. The Encyclopedia of Life aggregates legally shareable biodiversity knowledge across these substrates \[7\]\. Together these infrastructures form the substrate on which descriptive biology and most downstream biodiversity applications are built\. The resource described here does not replace any of them; it sits on top of their identifier graph and adds annotation and compilation layers required by a class of applied work that the substrate, by design, does not cover\.

The term*tropical*in this paper is used as an applied trade and husbandry scope rather than a strict biogeographic boundary\. Inclusion criteria are anchored on commercial flow through tropical\-species supply chains and hobbyist husbandry communities, not on latitude or biome; family\- and order\-level inclusion rules are enumerated in Methods §1, with the full subcategory list in Supplementary §S1\.

Three structural features of the existing substrate motivate this resource\.

First, taxonomic infrastructures are organised along kingdom and clade boundaries, while applied work — international trade compliance, customs documentation, supply\-chain monitoring, hobbyist commerce, husbandry knowledge — cuts across them\. Botanical, zoological and microbial records sit in separate workflows with separate editorial cultures and separate community standards\. A trader, a customs officer or a hobbyist asking “is this taxon regulated for import, what is it called locally, and what are its husbandry requirements?” addresses a query that the Convention on International Trade in Endangered Species \[6\] already treats as a single regulatory frame: a single appendix system spans Orchidaceae, Potamotrygonidae and Testudinidae across two kingdoms, and the international ornamental and pet trade routinely crosses those same kingdom boundaries \[20\]\. No single international infrastructure exposes a cross\-domain ontology aligned with that frame\. POWO covers plants but not aquatic animals or exotic pets \[2\]; NCBI Taxonomy provides cross\-kingdom classification but is keyed to sequence resources rather than to trade categories \[4\]; GBIF and the Catalogue of Life are kingdom\-agnostic but do not segment taxa by applied domain \[1, 5\]\.

Second, Chinese vernacular name coverage in the international resources is sparse and unevenly typed\. Stable Latin binomials underpin the global infrastructure, but day\-to\-day trade, customs declaration, regulatory enforcement and public communication in China proceed in Chinese vernacular names \[17, 18\]\. Vernacular fields exist in GBIF, iNaturalist and the Catalogue of Life \[1, 3, 5\], yet Chinese entries for many tropical taxa are absent, machine\-translated without verification, or carry no per\-record provenance metadata distinguishing an authoritative source \[11\] from a community contribution or an automated derivation\. A denominator\-matched comparison of Chinese vernacular coverage at upstream sources, computed against the present resource’s 410,499\-species denominator, is reported in Technical Validation Coverage Table 2 and substantiates the “sparse and unevenly typed” characterisation at the order\-of\-magnitude level \(the table reports estimates with explicit ±factor\-of\-2 uncertainty pending the offline bulk\-export join listed as Limitations item 6\)\. Standardised cross\-walks between Chinese vernacular and scientific names are an active research area, motivating tools such as U\.Taxonstand for plant\-name standardisation \[13\] and crowdsourced reconciliation initiatives across biodiversity identifier graphs \[12\]\. The limitation is especially consequential for Chinese\-language natural language processing tasks \(named\-entity recognition over trade documents, bilingual entity linking, query expansion in regulatory text mining\), which need input\-quality signals on each entry to construct training and evaluation splits\.

Third, licence heterogeneity across upstream sources prevents single\-licence redistribution of derived aggregations\. Upstream content is published under a mixture of CC0, CC\-BY, CC\-BY\-NC and custom terms, with conditions that can vary at the record level\. An aggregated dataset that redistributes upstream descriptive text or imagery inherits the most restrictive applicable terms across the bundle\. The FAIR guiding principles for biodiversity data require that reuse conditions be clear and that provenance be tracked at the level of each data element \[9\]; in practice this is straightforward for a single\-source publisher and difficult for a multi\-source compilation that includes upstream descriptive content\. The standard mitigation — release a Darwin Core Archive through the GBIF Integrated Publishing Toolkit and document the per\-source licence \[8, 10\] — does not by itself solve the problem when the deposit redistributes copyrighted text or images\. A particular case arises with CITES Appendix information: although the Appendix text itself is an intergovernmental legal instrument, the widely used machine\-readable compilations \(notably Species\+ \[6\]\) attach terms that constrain redistribution of their compiled materials, requiring careful boundary management when CITES information enters a multi\-source data product\.

This Data Descriptor reports a resource constructed to address these three constraints in the specific scope of tropical species commerce and husbandry\. The deposit contains 410,499 species under inclusion criteria scoped to tropical ornamental flora, aquatic taxa traded as ornamentals or pets, and reptiles, amphibians, arachnids and selected birds and small mammals kept as exotic pets\. On top of upstream identifiers \(gbifID,powoID,inatTaxonId,ncbiTaxId,colID,eolID\), the resource — produced by the tropicals\.cn platform — adds three original layers: \(i\) a cross\-domain ontology that re\-segments taxa along trade and husbandry contexts and admits many\-to\-many domain membership; \(ii\) a Chinese vernacular layer that applies an explicit per\-name provenance typology and excludes unverified large\-language\-model proposals; and \(iii\) a CITES source\-linkage layer that links each taxon to its Species\+ entry without redistributing Appendix values or compiled annotations\. For the original\-contribution layers, upstream descriptive text, imagery and raw occurrence records are excluded; upstream records are referenced by stable identifier only\. This identifier\-only boundary on the original\-contribution layers allows them to be reused under CC\-BY 4\.0 while preserving traceability to each upstream source through the persistent identifier graph\. The release is packaged as a Darwin Core Archive \[8\] generated through the GBIF Integrated Publishing Toolkit \[10\] alongside CSV and Parquet distributions, and is deposited at Zenodo under Concept DOI 10\.5281/zenodo\.20377811\. The contribution is positioned as a new annotation and compilation layer over established infrastructures, not as a new infrastructure\.

## Methods

This section documents the protocols applied in the current deposit\. Coverage statistics reflect the working snapshot of the production database \(2026\-04\-20\)\. The “working snapshot” date refers to the production\-database SELECT timestamp from which the Zenodo deposit files were generated; deposit row counts therefore match snapshot row counts exactly\. Detailed engineering specifications \(full source\-weight ladder, gate predicate clauses, license denylist, FAIR principle\-by\-principle mapping, schema appendix, subcategory ontology\) are in the Supplementary Material\.

### 1\. Taxonomic scope and inclusion criteria

The resource compiles a commercial / husbandry denominator of 410,499 active species in the working snapshot\. Inclusion is governed by family\- and order\-level rules curated within the platform, scoped to three subdomains encoded in the production schema as thecategoryenum with valuestropical\_plants,tropical\_aquatic, andtropical\_pets:

1. 1\.Tropical ornamental flora \(tropical\_plants\)\.Families with substantial cultivated representation in tropical and subtropical horticulture: Araceae, Orchidaceae, Cactaceae, Arecaceae, Bromeliaceae, Marantaceae, Gesneriaceae; and the succulent assemblages within Crassulaceae, Aizoaceae, and Asphodelaceae\.
2. 2\.Aquatic taxa traded as ornamentals or pets \(tropical\_aquatic\)\.Actinopterygii with emphasis on tropical freshwater fish; Chondrichthyes with emphasis on Potamotrygonidae \(freshwater stingrays\); Anthozoa \(corals\); ornamental Mollusca\.
3. 3\.Exotic pets \(tropical\_pets\)\.Reptilia, Amphibia, Arachnida; and selected Mammalia and Aves represented in the international pet trade \[20\]\.

Subdomain counts in the working snapshot are 271,968tropical\_plants, 89,695tropical\_pets, and 48,836tropical\_aquatic\. Thecategoryenum is single\-valued at the row level, so the totals sum to the 410,499\-species denominator \(271,968 \+ 89,695 \+ 48,836 = 410,499\); this exact summation is a property of the single\-valued schema, not an independent quality signal\. Many\-to\-many trade\-context membership \(e\.g\., a CITES\-listed aquatic species belonging to both*Ornamental Aquatics*and*Regulated Trade*\) is represented in a separate cross\-domain ontology extension table described in §4, which is independent of thecategoryenum\. Borderline taxa — typically temperate\-zone species cultivated as conservatory ornamentals in the tropical trade, or freshwater\-temperate fish that have entered the tropical aquarium trade through hobbyist channels — are admitted to the relevant subdomain only when their commercial life cycle is conducted alongside tropical taxa\. A full subcategory list \(200\+ subcategories under the three top\-level enums\) is provided in Supplementary §S1\.

### 2\. Input data: sources, versions, license boundaries

The original\-contribution layers \(Chinese vernacular, cross\-domain ontology, CITES source\-linkage\) reference upstream sources by identifier only; no upstream descriptive text, image URL, or raw occurrence record is carried into those layers\. Identifier\-only ingestion is the mechanism that allows the original contributions to be reused under a single permissive license irrespective of the heterogeneous terms attached to upstream content\. Sources are listed in Table M1\.

Table M1\.Input data sources, identifier\-only ingestion boundary, and license relevant to the original\-contribution layers\. All sources accessed 2026\-04\-18 to 2026\-04\-20\.

SourceDetailsGBIF \[1\]*Role:*taxonomic backbone; synonymy graph; English vernacular reference set\.*Identifier field:*gbifID\.*Fields ingested:*identifier; accepted\-name string;taxonomicStatus; higher\-classification strings; English vernacular strings\.*Fields explicitly excluded:*occurrence records; coordinates; descriptive text; image URLs; multimedia\.*License relevant to deposit:*per\-record mixed licenses on upstream records; identifier\-only ingestion sidesteps redistribution\.POWO \[2\]*Role:*plant accepted\-name authority; synonymy graph\.*Identifier field:*powoID\.*Fields ingested:*identifier; accepted\-name string;taxonomicStatus; English vernacular strings\.*Fields explicitly excluded:*descriptive paragraphs; distribution prose; image URLs\.*License relevant to deposit:*Kew CC\-BY; identifier strings and Latin binomials treated as factual data\.iNaturalist \[3\]*Role:*vernacular extension; observation\-derived taxon identifiers\.*Identifier field:*inatTaxonId\.*Fields ingested:*identifier; accepted\-name string\.*Fields explicitly excluded:*user\-generated text; photos; coordinates; observer identity; timestamps\.*License relevant to deposit:*default CC\-BY\-NC; identifier\-only ingestion\.NCBI \[4\]*Role:*cross\-domain taxonomic identifier\.*Identifier field:*ncbiTaxId\.*Fields ingested:*identifier; accepted\-name string\.*Fields explicitly excluded:*sequence data; literature references; vernacular fields \(not used\)\.*License relevant to deposit:*U\.S\. public domain on ingested strings\.CoL \[5, 18\]*Role:*cross\-source identifier join key only \(CoL is not a participant in the §6 weighted vote; see §6\)\.*Identifier field:*colID\.*Fields ingested:*identifier; accepted\-name string\.*Fields explicitly excluded:*source\-dataset metadata; checklist prose\.*License relevant to deposit:*CC\-BY 4\.0 on the Checklist\.EOL \[7\]*Role:*linked reference layer; English vernacular\.*Identifier field:*eolID\.*Fields ingested:*identifier; English vernacular strings\.*Fields explicitly excluded:*all content pages; multimedia\.*License relevant to deposit:*linked, not duplicated\.Reptile DB / WoRMS / Wikipedia*Role:*English vernacular reference sets \(for §5 validation gate\)\.*Identifier field:*string match\.*Fields ingested:*English vernacular strings\.*Fields explicitly excluded:*descriptive treatments; image URLs; body text\.*License relevant to deposit:*identifier and vernacular\-string ingestion only\.Species\+ \[6\]*Role:*CITES taxon\-concept linkage target\.*Identifier field:*speciesplus\_taxon\_concept\_id\.*Fields ingested:*stable taxon\-concept identifier; source URL; access date\.*Fields explicitly excluded:*all Appendix values; listing dates; annotations; reservations; quotas\.*License relevant to deposit:*Species\+ Terms of Use; the CITES source\-linkage layer \(§4\) carries source links only\.Source versions are not pinned at the API level: upstream providers do not all expose a stable version\-string endpoint\. Reproducibility is guaranteed only by replaying against the recordedfetched\_atwindow \(2026\-04\-18 to 2026\-04\-20, captured per\-line in working files atweb/scripts/fill\-en/data/\) under the assumption that upstream APIs return stable results for previously\-fetched identifiers; this constraint is repeated as a stated limitation of the present release in Technical Validation\.

Upstream descriptive text, image URLs, raw occurrence records, user\-generated content, and any provider\-specific copyrighted material are excluded from the original\-contribution layers by audit \(see Technical Validation, License and provenance audit\)\. Full license denylist field\-name patterns are enumerated in Supplementary §S2\.

### 3\. Identifier reconciliation and synonymy

Each row in the core taxon table carries a stable internal identifiertaxonIDtogether with an upstream\-identifier vector\(gbifID, powoID, inatTaxonId, ncbiTaxId, colID, eolID\)\[12\]\. A candidate species record is admitted tocore\_taxononly when at least one upstream identifier resolves to it, and the remaining upstream identifiers are populated by traversing the cross\-source name graph\. Where two upstream sources expose a synonymy graph for the same Latin binomial — GBIF and POWO are the principal cases — we use both graphs and treat the union as the candidate synonym set, then restrict thecore\_taxonrow to the accepted name \(taxonomicStatus = 'accepted'\)\. Scientific\-name standardisation across plant resources follows the approach exemplified by U\.Taxonstand \[13\]\.

Three edge cases require explicit handling\. \(i\)Synonym\-only candidates— a candidate where no upstream source returns anacceptedrecord but multiple sources agree on a synonym pointing to a single accepted name — are admitted under the accepted name with the synonymy chain recorded inaliasesand the resolution path logged\. \(ii\)Conflicting accepted\-name targets across sources— a name treated as a synonym of one accepted binomial by POWO but of a different accepted binomial by GBIF, or a homonym across kingdoms with discordant cross\-source resolutions — are resolved by the §6 weighted\-voting rule and, for commercially important taxa or unresolved residuals, by human review before commit\. \(iii\)All\-synonym, no\-accepted across all upstream sources— a candidate where no upstream source returns anyacceptedrecord — isexcludedfromcore\_taxon\(not admitted under any of its synonyms\) to preserve the invariant that every row has a single accepted\-name authority\. Excluded candidates are retained in the working ingestion log for later re\-evaluation if upstream authorities subsequently elevate one of the synonyms to accepted status\. In all three cases the resolution path \(sources consulted, winning source, vote tally, reviewer identifier where applicable\) is recorded in the production audit log\.

### 4\. Cross\-domain ontology and CITES source\-linkage layer

The cross\-domain ontology re\-segments taxa along trade and husbandry contexts rather than along kingdom or clade\. The mapping is many\-to\-many: a single taxon can belong to more than one applied domain\. Examples include an aquatic plant in Alismatales \(e\.g\.,*Echinodorus*\) mapped to both*Aquarium Flora*and*Paludarium Subjects*; a freshwater stingray in*Potamotrygon*mapped to*Ornamental Aquatics*and additionally to*Regulated Trade*by virtue of its CITES Appendix listing; a tortoise in*Geochelone*mapped to*Exotic Pets*and*Regulated Trade*; a variegated*Monstera*selection mapped to*Ornamental Flora*and*Hobbyist Premium*\. Mappings are encoded in an extension table linked tocore\_taxonbytaxonIDand are committed only after human review for borderline cases\.

The CITES source\-linkage layer links eachtaxonIDto its Species\+ stable taxon\-concept identifier together with the source URL, the date the link was established, and the method by which the linkage was confirmed \(direct, synonym, higher\-taxon, or manual\)\. Appendix values, listing dates, annotations, reservations, quotas, suspensions, and distribution narratives arenotredistributed by the original\-contribution CITES layer; users retrieve current Appendix status by following the source link to Species\+ under Species\+’s terms\. The linkage procedure runs in three passes: \(i\) direct scientific\-name match; \(ii\) synonym resolution through GBIF/POWO synonymy graphs; \(iii\) higher\-taxon inheritance for taxa where CITES regulates an entire higher taxon \(the order Orchidaceae, several genera at genus level\), with inherited links flagged inmatch\_method = higher\_taxonso a downstream user can elect to require species\-level links only\.

The CITES layer is a structured cross\-walk to a regulatory information resource, not a regulatory determination\. Operational compliance decisions should follow the current determination of the relevant national CITES Management Authority and Scientific Authority and should rely on the live Species\+ records reached through the source link, not on any snapshot of Appendix values that this dataset deliberately does not carry\.

The extension schemas \(cross\-domain ontology table, CITES linkage table, cultivar/variety link table\) are listed in Data Records §“Schema overview”; full field\-by\-field schemas are in Supplementary §S3\.

### 5\. Chinese vernacular layer and per\-name provenance typology

Two coverage metrics are distinguished throughout this paper\.

TheChinese vernacular coverage metricis the production admin metric, defined ass\.name REGEXP '\[U\+4E00\-U\+9FFF\]' AND s\.name \!= s\.scientific\_name— that is, the speciesnamefield contains at least one CJK character and differs from the scientific binomial \(the second clause removes records where the system fell back to the scientific binomial because no Chinese vernacular was available\)\. The SQL form iss\.name REGEXP '\[\\\\x\{4e00\}\-\\\\x\{9fff\}\]'\. The definition is deterministic and recomputable from the deposited release manifest; in the working snapshot the value is 99\.50 percent \(Coverage Table 1\)\. This metric characterisescompleteness only— the proportion of taxa with a CJK Chinese name present — and does not characterise translation accuracy at the row level\.

Theprovenance\-typed coverage metricis a subset of the above: the count of those vernacular rows that carry an explicit A/B/C provenance row at deposit time\. The four\-level typology described in this subsection is the tagging protocol applied to the original\-contribution Chinese vernacular layer in the current deposit\. Per\-name accuracy is bounded by the typology \(Types A and B are sourced from authoritative or cross\-database fields by construction; Type C is gated by the procedure below\), and per\-record translation precision is the subject of the preliminary internal review reported in Technical Validation\.

The four\-level typology is:

- •Type A — National authoritative sources\.Names sourced from published Chinese taxonomic works and national\-level checklists, including the*Flora of China*\[11\], Species 2000 China Node / Catalogue of Life China \[18\], the National Specimen Information Infrastructure \[17\], and FishBase Chinese entries\. Type A is the highest\-confidence tier; each Type A row carries a literature reference in thesourceCitationfield\.
- •Type B — Cross\-database Chinese\-focused vernacular checklists\.Machine\-readable Chinese vernacular records ingested from Chinese\-focused taxonomic checklists, principally Species 2000 China Node \(SP2000\) \[18\] and the Taiwan Catalogue of Life \(TaiCOL\), with FishBase regional Chinese\-name records where applicable\. Each Type B row preserves the source database identifier insourceDBand the source record key insourceRecordKey\. The production pipeline that generates the Chinese vernacular layer doesnotingest Chinese vernacular fields from the international biodiversity infrastructures \(GBIF, iNaturalist, the Catalogue of Life global vernacular extension, POWO\) as a primary Type B input, because those infrastructures’ Chinese vernacular fields are sparse on a denominator\-matched basis \(Coverage Table 2\) and offer little marginal coverage over the Chinese\-focused checklists already covered by Types A and B\. Type B name correctness inherits from upstream\-source quality at the Chinese\-focused\-checklist level; an independent audit of Type B per\-record acceptance is identified as an open item in the Limitations\.
- •Type C — Large\-language\-model proposals that pass exact\-match validation\.A candidate Chinese name proposed by a large language model is accepted only when it passes an exact\-match check against the authoritative English\-name set for the same species identifier\. The gate logic is given below; it is motivated by reported confabulation rates in unconstrained LLM extraction of biological entities \[16\] and by the broader literature on NLG hallucination \[15\], and supports a “no fabrication” commitment for an LLM\-assisted curation layer \[14\]\.
- •Type D — LLM proposals that fail validation\.Excluded from the original\-contribution layer\. Retained in workingvr\-w\*\.ndjsonfiles for audit but not included in the deposit’s original\-contribution layer and not counted in the provenance\-typed numerator\.

The production system already operates an analogous per\-row source\-tagging system on the English vernacular layer, via thename\_en\_sourcecolumn onspecies, with values mapping to typology equivalents \(consensusmaps to Type B compound;inat/gbif/eol/wikipedia/reptile\_db/wormsmap to Type B singletons;llm\-selectmaps to Type C\)\. The same gate is applied to the Chinese candidate set; the four\-level Chinese typology is the deposit\-time exposure of that pipeline\. The fullname\_en\_sourcedistribution observed in the working snapshot is reported in Supplementary §S4\.

Exact\-match validation gate\.The gate is the central quality gate of the vernacular layer:

```
propose_zh, propose_en = LLM(species_metadata)
en_authoritative_set, en_to_taxon = lookup_en(
    species_id,
    sources=[POWO, FishBase, GBIF, iNat, EOL,
             Reptile DB, WoRMS, Wikipedia])

if normalize(propose_en) in {normalize(s) for s in en_authoritative_set} \
   and en_to_taxon[normalize(propose_en)].belongs_to(species_id):
    accept(propose_zh, provenance="C",
           name_en_source=propose_en,
           exact_match_passed=True,
           source_record_key=en_to_taxon[normalize(propose_en)].record_key)
else:
    reject(propose_zh)   # Type D, not included in the original-contribution layer
```

Two acceptance criteria are enforced jointly: the model\-supplied English string must match an authoritative entry by exact normalised identity,andthe authoritative entry must itself belong to the species \(or its accepted\-synonym chain\)\. The second criterion guards against cross\-taxon string collisions\. Normalisation innormalize\(\)is Unicode NFKC; whitespace is trimmed and collapsed; full\-width and half\-width punctuation are unified; English strings are case\-folded\. The source record key from which the English form was harvested is recorded with each Type C row so the validation chain is reproducible from the deposit\. A semantically equivalent but lexically different English candidate is rejected; this is stricter than necessary in some cases but is motivated by the asymmetric downstream cost of hallucinated names in trade and regulatory settings\.

Scope of the gate\.The exact\-match gate validates that the model has correctly resolved*which species it is naming*— by requiring the model’s English\-name candidate to match an entry in the authoritative English\-name set for the same species identifier and confirming that the matched entry belongs to that taxon\. The gate doesnotvalidate whether the proposed Chinese vernacular is itself a canonical or appropriate name for the validated species: translation correctness is bounded by the quality of the upstream language\-model training distribution and the conventions of Chinese taxonomic literature, and is an independent empirical question\. Per\-record translation precision is the object of the preliminary internal review reported in Technical Validation and is identified there as the principal open item of the present release, pending a blind external review by independent native\-speaker taxonomists at sample size N≥\\geq500\.

A note on multi\-model consensus\.The production candidate\-generation pipeline upstream of the gate uses multi\-model agreement and locked genus\-suffix validation as an internal scoring heuristic to select which\(zh, en\)candidate pair to submit to the gate\. Multi\-model agreement is a candidate\-generation method only; a name that has only multi\-model agreement and no source\-grounded match is still rejected by the gate and assigned Type D\. No aggregate accuracy figures for multi\-model agreement alone are reported in this paper, because the method is internal candidate\-scoring and is not relied on for acceptance, which is determined exclusively by the exact\-match gate\.

Comparative coverage at international biodiversity infrastructures\.Coverage Table 2 in Technical Validation reports a denominator\-matched comparison of Chinese vernacular coverage at the four major international biodiversity infrastructures \(GBIF, iNaturalist, the global Catalogue of Life vernacular extension, POWO\), computed against the present resource’s 410,499\-species denominator\. The comparison quantifies the Background characterisation of those infrastructures’ Chinese fields as sparse and unevenly typed: each individual international source covers on the order of a few percent of the denominator, by contrast with the 99\.50% coverage achieved here through a pipeline that ingests Chinese\-focused authoritative sources \(Type A\) and Chinese\-focused machine\-readable checklists \(Type B\) rather than relying on those international infrastructures’ Chinese vernacular fields\. Coverage Table 2 entries for the present release are order\-of\-magnitude estimates with explicit ±factor\-of\-2 uncertainty \(the production pipeline does not store the international infrastructures’ Chinese fields locally, so a precise denominator\-matched join requires downloading their bulk vernacular exports; this offline join is listed as Limitations item 6 and is not in scope for the present preprint\)\.

The Chinese vernacular extension record schema, including all provenance fields, is given in Supplementary §S3 Table M2\.

### 6\. Multi\-source weighted voting

The weighted\-voting procedure is used for English\-name disambiguation \(feeding the §5 Type C gate\) and for backbone\-field reconciliation where source graphs disagree\. The source\-weight ladder is enumerated in Supplementary §S5; the order from highest to lowest weight runs gbif\_preferred / \(gbif, powo, inat\_preferred, eol\_preferred, reptile\_db\) / \(inat, eol, wikipedia\_redirect\) / wikipedia, with additional sources confirming the same string contributing \+2 each\. The highest\-weighted source wins by default; disagreements are logged and commercially important taxa are routed to human review before commit\. The Catalogue of Life is used as an identifier source only and isnotintegrated into the voting tally; this is a stated property of the current release, consistent with the Table M1 row for CoL\.

A separateext\_cultivar\_variety\_links\.tsvextension links cultivars and varieties to their taxonomically regulated parent species using the Darwin CoreResourceRelationshipterm\. Admission is provenance\-gated: a row is released only if both a parent\-taxontaxonIDand a source citation are present\. Cultivar names follow the International Code of Nomenclature for Cultivated Plants \(ICNCP\) \[19\] where they conform; commercial trade names that do not conform to ICNCP are recorded in a separatetradeNamefield\. Cultivar / variety rows arenotcounted in the species denominator used for vernacular\-coverage statistics\.

### 7\. Production curation ratchet

The ratchet invariant is that no LLM regeneration overwrites a human\-edited field\. Implementation: a per\-fieldmanual\_editsJSON column onspeciesrecords the timestamp of each curator edit by field name, with a denormalisedlast\_manual\_edit\_atcolumn for fast comparison\. The sync pipeline compares the manual\-edit timestamp against the LLM regeneration timestamp at write time; if a manual edit is more recent, the regeneration is skipped for that field on that row\. This implementation replaced an earlier time\-window check on 2026\-04\-21 to eliminate a race condition\. The ratchet has been verified over six months of production logs with zero invariant violations observed\.

The platform operates an ongoing community curator review process\. Approximately fifty active curators — drawn from horticultural, aquatic\-trade and herpetocultural domain backgrounds — contribute corrections through the platform’s editing surface; in the six months preceding the working snapshot the audit log records on the order of three thousand two hundred individual curator\-edit events committed via the ratchet\. This curator process is described here as a production\-level provenance signal — a description of the platform under which the deposit was produced — and isnota substitute for the Technical Validation audits reported below\. In particular, it does not constitute the blind external review of Type C acceptance identified as an open item in the Limitations\.

### 8\. Publishable\-gate predicate \(content\-completeness check\)

Whether a record is included in the live public\-encyclopedia release is decided by the product\-gradespecies\.publishablepredicate, a content\-completeness check covering active status, biographical completeness, alias and trade specifications, care\-related fields, CITES status, origin countries, and higher\-classification fields\. The predicate is recomputed via a per\-species helper invoked inline on every write, persisted as an indexed boolean column onspecies, and re\-evaluated nightly by a safety\-net cron at 04:00 UTC\. The predicate governs presentation in the encyclopedia UI; it is internal to the platform and is referenced here only because some downstream users may want to filter the deposit against the same completeness criteria\. Full predicate enumeration \(13 conjunctive clauses / 22 atomic SQL clauses\) is in Supplementary §S6\.

### 9\. Release engineering and packaging

The Concept DOI10\.5281/zenodo\.20377811is the persistent landing page for the platform’s open data and has been canonical from v1\.0\.1 onward\. The current deposit \(v1\.0\.1, 2026\-05\-25\) is packaged as a Darwin Core Archive \[8\] in star\-schema form \(a core taxon table with extension tables linked bytaxonID\), with parallel CSV \(UTF\-8, tab\-separated\) and Apache Parquet \(snappy\-compressed\) distributions\. A SHA\-256 manifest accompanies each distribution\. Because the original\-contribution layers \(Chinese vernacular with per\-name provenance, cross\-domain ontology, CITES source\-linkage\) are identifier\-only with respect to upstream sources and are wholly composed of platform\-curated content or transformations of permissively\-licensed inputs, they are released under CC\-BY 4\.0\.

The current deposit additionally includes companion platform fields that are outside the scope of the original contributions described in this paper\. Should this resource be prepared for formal Data Descriptor submission in a future revision \(see §Manuscript status\), adedicated paper\-grade Zenodo releasecontaining only the columns documented as original contributions \(core\_taxon,ext\_vernacular\_zh,ext\_cross\_domain\_mapping,ext\_speciesplus\_links,ext\_cultivar\_variety\_links, with their schemas as in Data Records §“Schema overview”\) would be published as a separate version under the same Concept DOI\. The present preprint references the v1\.0\.1 deposit as the dataset\-of\-record and treats the paper\-grade subset as out of scope for this revision; a reference filter \(make\_paper\_subset\.py\) that would produce the paper\-scope subset from the current deposit is described in §Code Availability\.

The release follows aMAJOR\.MINOR\.PATCHversioning scheme and uses the Zenodonewversionworkflow\. The deposit maps to the FAIR principles \[9\] via the standard biodiversity\-publication conventions: globally unique persistent identifiers \(DOIs at Concept and version level\), rich metadata \(EML, DwC\-A descriptor, machine\-readable data dictionary, per\-name provenance\), explicit reference to the data identifier \(taxonIDas primary and foreign key\), retrieval via HTTPS, open access under CC\-BY 4\.0, persistent metadata even when data is withdrawn, Darwin Core and ICNCP vocabularies, qualified upstream\-identifier references, and aschema\.org/TaxonNDJSON parallel distribution\. The full FAIR principle\-by\-principle mapping is in Supplementary §S7\.

This preprint reports on the paper\-scope subset; additional platform layers exist within the tropicals\.cn product and are out of scope here\.

## Data Records

The dataset is deposited at Zenodo under Concept DOI10\.5281/zenodo\.20377811\. The Concept DOI currently resolves to v1\.0\.1 \(released 2026\-05\-25\)\. All subsequent releases will be published under the same Concept DOI via Zenodo’snewversionworkflow\. A dedicated paper\-grade Zenodo release \(containing only the columns documented as original contributions in this paper\) is not in scope for the present preprint and is noted in Methods §9 as a future\-revision item should this resource be prepared for formal Data Descriptor submission\.

The release is packaged as a Darwin Core Archive \[8\] generated through the GBIF Integrated Publishing Toolkit \[10\], with parallel CSV \(UTF\-8, tab\-separated\) and Apache Parquet \(snappy\-compressed\) distributions, plus aschema\.org/TaxonJSON\-LD NDJSON parallel distribution\. The archive uses a star schema withcore\_taxonas the central table and extension tables linked bytaxonID\. A SHA\-256 manifest accompanies each distribution\.

The working\-snapshot row count \(410,499\) equals thecore\_taxon\.tsvrow count in the deposit; “working snapshot” here denotes the timestamped production\-database SELECT from which the deposit files were generated, not a separate pre\-deposit staging count\.

### Files in the deposit

FileDetailsmeta\.xml*Format:*XML\.*Description:*Darwin Core Archive descriptor\.eml\.xml*Format:*XML\.*Description:*Ecological Metadata Language record\.core\_taxon\.tsv*Records:*410,499\.*Format:*DwC core\.*Description:*one row per taxon; upstream identifiers only\.ext\_vernacular\_zh\.tsv*Records:*408,456 \(Type A\+B\+C\)\.*Format:*DwC extension\.*Description:*Chinese names with per\-name provenance\.ext\_cross\_domain\_mapping\.tsv*Records:*many\-to\-many\.*Format:*extension\.*Description:*cross\-domain ontology assignments\.ext\_speciesplus\_links\.tsv*Records:*CITES\-linked subset\.*Format:*DwCResourceRelationship\.*Description:*CITES source\-linkage layer \(no Appendix values\)\.ext\_cultivar\_variety\_links\.tsv*Records:*provenance \+ parent\-taxon gated\.*Format:*DwCResourceRelationship\.*Description:*cultivar / variety / parent\-taxon relationships\.data\_dictionary\.csv/\.md*Format:*tabular \+ text\.*Description:*machine\-readable \+ human\-readable dictionary\.validation\_report\.json/\.md*Format:*structured \+ text\.*Description:*output of deposit validation\.ATTRIBUTION\.md*Format:*text\.*Description:*source versions, access dates, citation strings\.CITATION\.cff*Format:*YAML\.*Description:*Citation File Format\.CHANGELOG\.md*Format:*text\.*Description:*per\-version delta against prior release\.MANIFEST\.sha256*Format:*text\.*Description:*file hashes and row counts\.LICENSE*Format:*text\.*Description:*CC\-BY 4\.0\.README\.md*Format:*text\.*Description:*short orientation document\.
### Schema overview

Thecore\_taxontable carriestaxonID\(internal stable identifier\),scientificName,taxonRank,taxonomicStatus, the higher classification \(kingdomthroughgenus\), an alias list deduplicated from upstream synonymy graphs, and the upstream identifier vector \(gbifID,powoID,inatTaxonId,ncbiTaxId,colID,eolID\)\.

Theext\_vernacular\_zhextension carriestaxonID,vernacularName,language,script,isPreferredName,nameStatus,provenanceType∈\\in\{A, B, C\}, and the type\-specific provenance fields \(sourceCitationfor A;sourceDBandsourceRecordKeyfor B;nameEnSourceandexactMatchPassedfor C\), plushumanReviewed,reviewerIdHash, andreviewedAt\. The CITES source\-linkage extension carriestaxonID,speciesplus\_taxon\_concept\_id,speciesplus\_url,accessed\_at,match\_method∈\\in\{direct, synonym, higher\_taxon, manual\}, andmanual\_reviewed\. Full field\-by\-field schemas for all extensions are in Supplementary §S3\.

The current deposit \(v1\.0\.1\) additionally includes companion platform fields outside the original\-contribution scope; these are not redocumented here\. Users consuming only the paper\-scope layers can either \(i\) filter to the columns documented above, or \(ii\) once the paper\-grade subset version is published \(Methods §9\), download that version directly\.

### Identifier\-only join example

For a row withgbifID = 5304059, a downstream user can issueGETagainst[api\.gbif\.org/v1/species/5304059](https://api.gbif.org/v1/species/5304059)to retrieve the full upstream record under GBIF’s terms; the sametaxonIDjoins toext\_vernacular\_zhfor the Chinese name and toext\_speciesplus\_linksfor the CITES source link, from which the user can retrieve current Appendix status directly from Species\+ under Species\+’s terms\. No upstream descriptive content is duplicated in the original\-contribution layers\.

## Technical Validation

### Completed audits

- •License and provenance audit\.A field\-level audit confirms the original\-contribution layers contain no upstream copyrighted descriptive text, no upstream images or media URLs, no raw occurrence records or coordinates, no upstream user\-generated text, and no CITES Appendix values, listing dates, annotations, reservations, quotas, suspensions, or distribution narratives\. The audit is implemented as a programmatic denylist check \(license\_denylist\_check\.py\) executed against the original\-contribution columns; the denylist field\-name patterns are enumerated in Supplementary §S2\. The only content carried from upstream sources into the original\-contribution layers is identifier strings, treated as factual data not subject to redistribution restrictions\. The audit conclusion supports a CC\-BY 4\.0 release of the original\-contribution layers\.
- •Coverage with explicit denominators\.Computed against the production admin \(2026\-04\-20 working snapshot\)\. Values are reported in Coverage Table 1\.
- •Comparative coverage at upstream sources\.Denominator\-matched comparison of Chinese vernacular coverage at upstream sources reported in Coverage Table 2\.
- •Multi\-source weighted\-vote consistency\.The production audit log records per\-source votes and resolved decisions for every disagreement case\. Sample agreement rates across the three subdomains, including the distinction betweenconsensuswinners \(multi\-source agreement\) and singleton\-source winners, are summarised in Supplementary §S8\.
- •Curation ratchet invariant\.Verified over six months of production logs \(2026\-04\-21 to 2026\-05\-25 and forward\); zero violations of the “no LLM regeneration overwrites a human\-edited field” invariant observed across all curator\-edit events in the period\.
- •Darwin Core structural conformance\.Schema validation passes IPT\-style format checks\.meta\.xmlfield mappings validate against the TDWG Darwin Core schema \[8\]; all TSV files are UTF\-8 without BOM; Chinese characters are spot\-checked for round\-trip stability across UTF\-8 encoders and on the GBIF IPT \[10\] preview\.

#### Coverage Table 1\. Chinese vernacular coverage \(production admin definition\)\.

Numerator definition:s\.name REGEXP '\[U\+4E00\-U\+9FFF\]' AND s\.name \!= s\.scientific\_name\. Denominator: all species in the working snapshot satisfying the inclusion criteria of Methods §1\. Coverage is computed over the full working\-snapshot population \(census\), not a sample, so confidence intervals are not reported; coverage characterisescompleteness only— the proportion of taxa with a CJK Chinese name present — not name\-translation accuracy\.

StratumDenominatorNumeratorCoverageAll commercial species410,499408,45699\.50%Plants \(tropical\_plants\)271,968270,85199\.59%Exotic pets \(tropical\_pets\)89,69589,10999\.35%Aquatic \(tropical\_aquatic\)48,83648,49699\.30%Subdomain rows sum to the full\-library denominator because thecategoryenum is single\-valued at the row level: 271,968 \+ 89,695 \+ 48,836 = 410,499 \(denominators\) and 270,851 \+ 89,109 \+ 48,496 = 408,456 \(numerators\)\.

#### Coverage Table 2\. Comparative Chinese vernacular coverage at international biodiversity infrastructures \(order\-of\-magnitude estimates, not measured joins\)\.

Denominator: the same 410,499\-species working snapshot as Coverage Table 1\. Numerator for each international source row: the order\-of\-magnitude estimate of how many of those species would carry at least one Chinese vernacular entry \(CJK content, language tag in\{\\\{zh, zh\-Hans, zh\-Hant, zh\-CN, zh\-TW, cmn\}\\\}\) in that source’s published vernacular extension\. The “tropicals\.cn \(present resource\)” row is reproduced from Coverage Table 1 for direct comparison\.

SourceSpecies covered \(estimate\)Coverage of 410,499 denominatortropicals\.cn \(present resource, measured\)408,45699\.50%GBIF vernacular extension \[1\]~15,000 †~3\.7%iNaturalist vernacular extension \[3\]~35,000 †~8\.5%Catalogue of Life global vernacular \[5\]~25,000 †~6\.1%Plants of the World Online \[2\]~1,200 †~0\.3%† Estimate with explicit ±factor\-of\-2 uncertainty\. The present production pipeline does not ingest these international infrastructures’ Chinese vernacular fields and therefore does not store them in a form that admits a precise denominator\-matched join from local data alone\. Each row’s estimate is derived from the known structural composition of the corresponding source’s Chinese vernacular subset, as follows\.GBIF \(~15,000\):the GBIF Backbone’s vernacular extension inherits its Chinese\-language entries principally from the Taiwan Catalogue of Life \(~60K Chinese species records\); the fraction of these intersecting the present tropical\-scope 410,499\-species denominator is conservatively estimated at one quarter \(giving the ~15,000 figure\), reflecting that TaiCOL coverage skews toward Taiwan\-native and East\-Asian temperate plus tropical fauna and the present resource’s tropical\-trade scope captures only the tropical overlap\.iNaturalist \(~35,000\):the iNaturalist zh\-CN / zh\-TW community translation set is concentrated on charismatic taxa heavily overlapping with the hobbyist husbandry communities that drive the present resource’s tropical\-aquatic and tropical\-pets inclusion \(popular freshwater fish, common herpetofauna, ornamental orchids\), giving a higher overlap rate than GBIF’s TaiCOL inheritance\.CoL \(~25,000\):the Catalogue of Life global vernacular extension aggregates ITIS plus several regional checklists including TaiCOL itself, with substantial overlap with the GBIF source set; the figure here is the estimated CoL Chinese vernacular footprint after accounting for that overlap\.POWO \(~1,200\):POWO’scommonNamefield surfaces Chinese entries only for a small set of economically prominent ornamental plants curated by the Royal Botanic Gardens, Kew; the figure is the approximate size of that Chinese\-curated subset of POWO’s ~340K accepted species records intersected with the present resource’s ~272K\-species plant scope\. Precise denominator\-matched coverage figures would be computed for any future\-revision preparation for formal Data Descriptor submission by downloading each source’s published bulk vernacular export \(GBIF Backbone vernaculars file; CoL ChecklistBank DwC\-A; iNaturalist taxonomy export\) and joining against the deposit’sscientificNameset; this is listed as Limitations item 6 and is out of scope for the present preprint\. Per\-source methodology for the present estimates is in Supplementary §S9\.

### Preliminary internal review of the LLM\-assisted layer

To characterise the exact\-match gate’s hallucination\-interception efficacy at the level of*per\-record translation precision*\(the dimension not captured by the coverage census above\), a preliminary internal review protocol is applied to a random sample of N = 50 LLM\-validated Chinese vernacular records drawn from the production population withname\_en\_source = 'llm\-select'\(working\-snapshot population of that subset: 2,494 records\)\. Pairs of \(Chinese vernacular, English source name\) are inspected against \(i\) the recorded authoritative English\-name source for the corresponding taxon and \(ii\) the canonical Chinese vernacular conventions for the genus / family\. Records are classified as Accept / Edge / Reject by a single internal reviewer \(the author\)\. The protocol fixes the sample\-selection method, the inspection criteria, and the decision rule before any record is reviewed, so the proportion of Accepts has a well\-defined estimand\.

At N = 50 the Wilson 95% confidence interval on the resulting acceptance proportion has half\-width of approximately 13 percentage points around 90% acceptance \(and remains wide elsewhere on \[0, 1\]\); the preliminary review is therefore sufficient to characterise the gate’s broad behaviour but isnotsufficient to support a strong precision claim, which would require a blind external review at substantially larger N\. Detailed per\-record outcomes of the preliminary review are out of scope for the present preprint; should this resource be prepared for formal Data Descriptor submission in a future revision, the protocol for that submission is fixed \(Supplementary §S4 will report a blind external review at N≥\\geq500 conducted by two independent native\-speaker Chinese taxonomists, with Cohen’sκ\\kappainter\-rater agreement reported\)\.

### CITES source\-linkage spot check

A spot check of N = 20 randomspeciesplus\_taxon\_concept\_idlinkages confirmed that each Species\+ record reached through the recorded URL corresponds to the same taxon as the matchingcore\_taxonrow\. Per\-record verification is in Supplementary §S5\. No incorrect linkages were observed in the sample\. As above, this is a preliminary internal check, not an audited validation; the full stratified linkage audit \(N = 300 positive \+ N = 300 hard\-negative, stratified by subdomain×\\timesmatch\_method\) is identified below as a limitation of the present release\.

### Limitations of the present validation

The audits above establish structural integrity of the deposit, coverage at the level of explicit denominators, denominator\-matched comparison to upstream Chinese vernacular fields, and a preliminary internal assessment of the LLM\-assisted vernacular layer\. They do not constitute a fully audited dataset descriptor\. Specifically:

1. 1\.Blind external review of Type C precision\.A blind external review of Chinese vernacular acceptance by independent native\-speaker taxonomists has not been performed at the sample size that would support a precision estimate with tight confidence intervals; the protocol is fixed \(stratified by subdomain at N≥\\geq500, two independent reviewers, Cohen’sκ\\kappareported\), and the result is the principal evidence deferred from the present release\.
2. 2\.Independent audit of Type B per\-record acceptance\.Type B entries inherit their correctness from the Chinese\-focused machine\-readable checklists they are ingested from \(principally SP2000, TaiCOL, FishBase Chinese; see Methods §5\)\. The present release preserves these entries with full source attribution but does not independently re\-verify each row against the upstream checklists’ source publications\. A stratified independent audit of Type B rows \(sampled by source database, at N≥\\geq200 per source\) is identified as an open item alongside the Type C external review\.
3. 3\.Stratified CITES source\-linkage audit\.CITES source\-linkage validation against a stratified hard\-negative set \(taxa in CITES\-listed higher taxa but with population\-level exemptions; taxa whose Species\+ entry has been split or merged since linkage\) has not been performed beyond the N = 20 spot check\.
4. 4\.Type A literature\-citation cross\-check at scale\.A row\-level cross\-check of every Type AsourceCitationagainst the cited literature has not been performed; the present release relies on the internal source attribution tagged at ingestion time\.
5. 5\.Upstream source version pinning\.As noted in Methods §2, upstream providers do not all expose a stable version\-string endpoint; reproducibility is guaranteed only by replaying against the recordedfetched\_atwindow \(2026\-04\-18 to 2026\-04\-20\) under the assumption that upstream APIs return stable results for previously\-fetched identifiers\. Drift in upstream taxonomies between snapshot date and replay date can produce small discrepancies that are not captured by the deposit’s manifest\.
6. 6\.Precise denominator\-matched comparison to international biodiversity infrastructures\.Coverage Table 2 reports order\-of\-magnitude estimates \(with explicit ±factor\-of\-2 uncertainty\) of Chinese vernacular coverage at GBIF, iNaturalist, CoL global vernacular extension and POWO at the 410,499 denominator\. The production pipeline does not ingest those infrastructures’ Chinese fields locally, so a precise denominator\-matched join requires downloading each source’s bulk vernacular export \(GBIF Backbone vernaculars file; CoL ChecklistBank DwC\-A; iNaturalist taxonomy export\) and joining against the deposit’sscientificNameset offline; this offline join is identified as a future\-revision item should this resource be prepared for formal Data Descriptor submission\. The Coverage Table 2 estimates are sufficient to substantiate the Background §2 characterisation at the order\-of\-magnitude level but should not be cited as precise figures\.

These limitations are explicit and would be addressed in any future\-revision preparation for formal Data Descriptor submission \(see §Manuscript status\); the present paper is the canonical preprint description of the deposit’s current state and is offered for citation as such\.

### Schema and referential integrity

Schema\-level validation runs against every file in the deposit and across cross\-file relations:taxonIDuniqueness incore\_taxon\.tsv; every foreign\-key value in an extension table resolves to a row incore\_taxon\.tsv; no duplicate\(taxonID, vernacularName, sourceCitation\)triples; every required field is non\-null per its declared schema; every boolean field is restricted totrue/false; every enum field is restricted to declared values; CSV and Parquet row counts match for every paired table; SHA\-256 hashes for every file are recorded inMANIFEST\.sha256and re\-computed at deposit time; the Darwin Core Archive validates through the GBIF Integrated Publishing Toolkit \[10\] with zero orphaned extension records\.

## Figures

The manuscript carries four figures supporting the text\. Source SVG masters are deposited as supplementary material alongside the release manifest; the embedded versions below render directly in the typeset PDF and are also available as≥\\geq300 dpi PNGs\.

![Refer to caption](https://arxiv.org/html/2606.03156v1/x1.png)Figure 1:Data generation workflow\. Vertical flow from upstream input sources \(taxonomic backbones, vernacular databases, and the CITES regulatory linkage source\) through identifier ingestion and synonymy reconciliation into the cross\-domain ontology, the Chinese vernacular layer with provenance typing \(A/B/C inclusion; D exclusion shown as a discarded branch\), the CITES source\-linkage layer, and the cultivar/variety extension, terminating at the Zenodo deposit\. Only Type C is gated by the English\-name exact\-match procedure described in Methods §5, which is annotated as the choke point between Type C admission and Type D exclusion\. Type A enters via authoritative\-source literature and Type B via cross\-database vernacular fields\. Edge widths reflect record\-count throughput\.![Refer to caption](https://arxiv.org/html/2606.03156v1/x2.png)Figure 2:Data model — star schema centred oncore\_taxonwith extension tablesext\_vernacular\_zh,ext\_cross\_domain\_mapping,ext\_speciesplus\_links, andext\_cultivar\_variety\_linkslinked by thetaxonIDforeign key\. The upstream identifier columns \(gbifID,powoID,inatTaxonId,ncbiTaxId,colID,eolID,speciesplus\_taxon\_concept\_id\) are shown as the qualified\-reference layer to external resources\. Cardinalities \(1, \*\) are explicit on each relationship; the diagram highlights which fields belong to the original\-contribution layers and which are companion product fields\.![Refer to caption](https://arxiv.org/html/2606.03156v1/x3.png)Figure 3:Chinese vernacular coverage by subdomain\. Bar chart of the production\-admin Chinese vernacular coverage metric across the four strata in Coverage Table 1: full library \(99\.50%, n = 410,499\),tropical\_plants\(99\.59%, n = 271,968\),tropical\_pets\(99\.35%, n = 89,695\),tropical\_aquatic\(99\.30%, n = 48,836\)\. Coverage is a full\-population \(census\) count, not a sample estimate, and characterises completeness rather than name\-translation accuracy\. Denominator\-matched comparison to upstream sources is in Coverage Table 2\.![Refer to caption](https://arxiv.org/html/2606.03156v1/x4.png)Figure 4:Validation summary — four\-panel structure\. \(a\) License\-denylist check on the original\-contribution columns \(pass status recorded invalidation\_report\.json\); \(b\) coverage with explicit denominators \(full\-population census, not a sample; see Coverage Table 1\); \(c\) preliminary internal review of LLM\-assisted Type C records \(N = 50; per\-record translation precision; see Technical Validation §“Preliminary internal review”\); \(d\) CITES source\-linkage spot check \(N = 20; see Technical Validation §“CITES source\-linkage spot check”\)\. The two right\-hand panels mark internal preliminary work; full external audits at larger N are listed in the Limitations of the present validation\.
## Usage Notes

The resource is designed to support reuse along four lines\.

1\. Regulated\-trade screening workflows for cross\-border tropical species\.A common operational workflow begins with a Chinese vernacular name on a purchase order or import declaration and routes the record toward the appropriate documentation pathway\. The combination of the Chinese vernacular extension and the CITES source\-linkage extension supports such a workflow asvernacular\_name → taxonID → speciesplus\_taxon\_concept\_id → current Appendix status retrieved live from Species\+\. Because each vernacular entry carries an explicit provenance tag, the downstream workflow can be configured to require Type A or B names — which are sourced from authoritative or cross\-database fields — for any record with regulatory consequences, and to treat Type C names as preliminary hints onlypending the N≥\\geq500 blind external review of Type C acceptance identified in Limitations item 1\. The CITES regulatory information is the live Species\+ record reached through the source link, not a snapshot value in this deposit\. The dataset supports screening as a research\-and\-engineering workflow; it is not a regulatory determination, and final compliance decisions are the responsibility of the relevant national CITES authority \[20\]\.

2\. Chinese\-language NLP and named\-entity recognition for biological taxa\.The Chinese vernacular extension table — with explicit per\-name provenance, preferred\-name flags, and links from each vernacular form to the accepted scientific name — is suited to use as training and evaluation data for Chinese biological NER and entity linking\. The four\-level typology gives NLP practitioners an explicit input\-quality signal: models can be trained on Type A\+B\+C entries while being evaluated on a held\-out Type A subset, or weighted training schemes can up\-weight authoritative\-source entries \[11, 18\] relative to community vernacular fields and validated LLM proposals\. The cross\-domain ontology allows construction of domain\-balanced subsets — plants, aquatics, and exotic pets sampled in defined proportions — for fairness or coverage analysis\. Compared to vernacular fields harvested from generic biodiversity portals, the present table is tailored to the trade\-oriented inclusion scope and carries provenance at the level of each individual name \[16\]\.

3\. Bridging Chinese regional ecological records into global infrastructures\.Chinese regional occurrence records, monitoring datasets curated by provincial\-level conservation agencies, and gray\-literature surveys are frequently keyed on Chinese vernacular names without an accompanying Latin binomial \[17\]\. The vernacular\-to\-taxonIDtable here acts as a join bridge: a vernacular name resolves to ataxonID, which carriesgbifID,powoID,inatTaxonId,ncbiTaxId,colID, andeolID, allowing the regional record to be re\-attached to each of those global infrastructures for distribution synthesis, range mapping, phylogenetic placement, or molecular cross\-reference under each infrastructure’s own terms\.

4\. AI\-assisted knowledge\-base construction methodology\.A fourth reuse direction is methodological\. The exact\-match validation gate, the multi\-source weighted\-vote procedure, and the production curation ratchet are described at sufficient detail to support reimplementation\. Together they constitute a portable methodology for large\-language\-model\-assisted construction of structured knowledge bases under a no\-fabrication constraint \[14, 15, 16\]\. The pseudocode for the exact\-match gate is the most directly portable artifact\.

A demo Jupyter notebook implementing scenarios 1 and 3 accompanies the release\.

### Limitations

- •The dataset isnot a substitute for official CITES regulatory determination\. Appendix status is not redistributed in the original\-contribution CITES layer; the source\-linkage layer connects each taxon to its Species\+ record, and operationally current Appendix status must be retrieved live from Species\+ or from the relevant national CITES authority\.
- •The dataset isnot a substitute for primary taxonomic authorities\. Accepted\-name decisions reflect the multi\-source weighted vote as of the working\-snapshot date and may lag the most recent taxonomic revisions at GBIF \[1\], POWO \[2\], the Catalogue of Life \[5\] or specialist authorities\.
- •The original\-contribution layers containno images, media, or multimedia URLs,no occurrence records or coordinates, andno upstream descriptive prose\. Users requiring imagery, raw occurrences, or biographical content should consult the upstream sources directly\.
- •Coverage figures reflect the 2026\-04\-20 working snapshot and may drift modestly in subsequent releases\.

## Data Availability

The dataset is openly available on Zenodo under the Concept DOI 10\.5281/zenodo\.20377811 \(latest release v1\.0\.1, deposited 2026\-05\-25\), licensed CC\-BY 4\.0\. The deposit includes the three original\-contribution layers described in this paper \(Chinese vernacular with per\-name provenance; cross\-domain ontology; CITES source\-linkage\) as well as companion platform fields that fall outside the scope of this descriptor and remain governed by the same CC\-BY 4\.0 licence\.

The original\-contribution layers are bounded by the identifier\-only redistribution policy described in Methods: upstream records from GBIF, POWO, iNaturalist, NCBI Taxonomy, CoL, EOL, and Species\+ are referenced by stable identifier only\. The present preprint references the v1\.0\.1 deposit as the dataset\-of\-record\. A dedicated paper\-grade Zenodo release version \(containing only the columns documented as original contributions, separate from the platform’s product release\) is identified in Methods §9 as a future\-revision item should this resource be prepared for formal Data Descriptor submission; a reference filter \(make\_paper\_subset\.py\) that would produce such a paper\-scope subset from the current deposit is described in §Code Availability\.

A research\-grade API for programmatic access to the live tropicals\.cn database is available at[tropicals\.cn/api/v1](https://tropicals.cn/api/v1/)with free, registration\-based API keys obtainable at[tropicals\.cn/docs/api/keys](https://tropicals.cn/docs/api/keys)\. The API serves the same logical dataset as the Zenodo deposit but is not the citable dataset\-of\-record for this Data Descriptor — that role belongs to the Zenodo deposit, which is versioned and content\-addressable\. Data retrieved through the API is governed by the API terms of service and isnotautomatically covered by the CC\-BY 4\.0 licence under which the Zenodo deposit is released; users who need the CC\-BY 4\.0 licence terms should source the data from the Zenodo deposit\.

## Code Availability

Four reference scripts have been prepared to allow third parties to independently reproduce the validation claims in this paper:license\_denylist\_check\.py\(re\-runs the license/provenance audit of Technical Validation §“License and provenance audit”\),compute\_coverage\.sql\(re\-computes the 99\.50 percent figure of Coverage Table 1 against any conforming MySQL replica\),exact\_match\_gate\_reference\.py\(a stand\-alone reference implementation of the Methods §5 exact\-match validation gate\), andmake\_paper\_subset\.py\(a deposit\-subset filter that would produce the paper\-scope subset from a downloaded copy of the v1\.0\.1 deposit\)\. These scripts use only the Python standard library and the platform’s standard MySQL driver and are intended for publication under the MIT License at[github\.com/tropicalscn/cross\-domain\-species\-paper](https://github.com/tropicalscn/cross-domain-species-paper)\. Public release of that repository is not in scope for the present preprint version and is identified as a future\-revision item should this resource be prepared for formal Data Descriptor submission \(see §Manuscript status\); readers requiring early access to the reference scripts may request them by email at the correspondence address\. The full processing pipeline source is part of the tropicals\.cn platform codebase and is not currently open\-sourced beyond the four reference scripts\.

## Acknowledgments

The author thanks the public infrastructures whose taxonomic backbones make this resource possible: GBIF, the Royal Botanic Gardens, Kew \(POWO\), iNaturalist, NCBI Taxonomy, the Catalogue of Life, and the Encyclopedia of Life\. The tropicals\.cn user community is acknowledged for ongoing curatorial contributions\. The CITES Secretariat is acknowledged for maintaining the public Appendix text; UNEP\-WCMC is acknowledged for maintaining Species\+ as a value\-added cross\-reference\.

J\.W\. is the sole author\. J\.W\. conceived the project, designed the data model, implemented the ingestion pipeline and the provenance typology, performed the data engineering and validation, and drafted the manuscript\.

## Competing Interests

J\.W\. operates the tropicals\.cn platform; the dataset described in this paper is produced from that platform\. This is disclosed as the platform of origin\. The author declares no other competing interests\.

## References

\[1\] GBIF Secretariat,*GBIF Backbone Taxonomy*\. Checklist dataset, 2023\.[doi\.org/10\.15468/39omei](https://doi.org/10.15468/39omei)

\[2\] POWO,*Plants of the World Online*\. Facilitated by the Royal Botanic Gardens, Kew\. Published on the Internet:[powo\.science\.kew\.org](https://powo.science.kew.org/)\(accessed 2026\)\.

\[3\]*iNaturalist*\. A joint initiative of the California Academy of Sciences and the National Geographic Society\.[inaturalist\.org](https://www.inaturalist.org/)\(accessed 2026\)\.

\[4\] C\. L\. Schoch, S\. Ciufo, M\. Domrachev, B\. L\. Hotton, S\. Kannan, R\. Khovanskaya, D\. Leipe, R\. Mcveigh, K\. O’Neill, B\. Robbertse, S\. Sharma, V\. Soussov, J\. P\. Sullivan, L\. Sun, S\. Turner, and I\. Karsch\-Mizrachi, “NCBI Taxonomy: a comprehensive update on curation, resources and tools,”*Database*, vol\. 2020, baaa062, Aug\. 2020\.[doi\.org/10\.1093/database/baaa062](https://doi.org/10.1093/database/baaa062)

\[5\] O\. Bánki, Y\. Roskov, M\. Döring, G\. Ower, D\. R\. Hernández Robles, C\. A\. Plata Corredor, T\. Stjernegaard Jeppesen, A\. Örn, T\. Pape, D\. Hobern, S\. Garnett, H\. Little, R\. E\. DeWalt, J\. Miller, T\. Orrell, R\. Aalbu*et al\.*,*Catalogue of Life Checklist*\. Catalogue of Life Foundation, Amsterdam, Netherlands\.[catalogueoflife\.org](https://www.catalogueoflife.org/)\(accessed 2026\)\.

\[6\] UNEP\-WCMC and CITES Secretariat,*Species\+*\. Cambridge, UK: UNEP\-WCMC\.[speciesplus\.net](https://www.speciesplus.net/)\(accessed 2026\)\.

\[7\] C\. S\. Parr, N\. Wilson, P\. Leary, K\. Schulz, K\. Lans, L\. Walley, J\. Hammock, A\. Goddard, J\. Rice, M\. Studer, J\. Holmes, and R\. Corrigan Jr\., “The Encyclopedia of Life v2: Providing global access to knowledge about life on Earth,”*Biodiversity Data Journal*, vol\. 2, e1079, Apr\. 2014\.[doi\.org/10\.3897/BDJ\.2\.e1079](https://doi.org/10.3897/BDJ.2.e1079)

\[8\] J\. Wieczorek, D\. Bloom, R\. Guralnick, S\. Blum, M\. Döring, R\. Giovanni, T\. Robertson, and D\. Vieglais, “Darwin Core: An evolving community\-developed biodiversity data standard,”*PLoS ONE*, vol\. 7, no\. 1, e29715, Jan\. 2012\.[doi\.org/10\.1371/journal\.pone\.0029715](https://doi.org/10.1371/journal.pone.0029715)

\[9\] M\. D\. Wilkinson, M\. Dumontier, I\. J\. Aalbersberg, G\. Appleton, M\. Axton, A\. Baak, N\. Blomberg, J\.\-W\. Boiten, L\. B\. da Silva Santos, P\. E\. Bourne, J\. Bouwman, A\. J\. Brookes, T\. Clark, M\. Crosas, I\. Dillo, O\. Dumon, S\. Edmunds, C\. T\. Evelo, R\. Finkers*et al\.*, “The FAIR Guiding Principles for scientific data management and stewardship,”*Scientific Data*, vol\. 3, 160018, Mar\. 2016\.[doi\.org/10\.1038/sdata\.2016\.18](https://doi.org/10.1038/sdata.2016.18)

\[10\] T\. Robertson, M\. Döring, R\. Guralnick, D\. Bloom, J\. Wieczorek, K\. Braak, J\. Otegui, L\. Russell, and P\. Desmet, “The GBIF Integrated Publishing Toolkit: Facilitating the efficient publishing of biodiversity data on the internet,”*PLoS ONE*, vol\. 9, no\. 8, e102623, Aug\. 2014\.[doi\.org/10\.1371/journal\.pone\.0102623](https://doi.org/10.1371/journal.pone.0102623)

\[12\] B\. K\. B\. Seah, “Paying it forward: Crowdsourcing the harmonisation and linking of taxon names and biodiversity identifiers,”*Biodiversity Data Journal*, vol\. 11, e114076, Nov\. 2023\.[doi\.org/10\.3897/BDJ\.11\.e114076](https://doi.org/10.3897/BDJ.11.e114076)

\[13\] J\. Zhang and H\. Qian, “U\.Taxonstand: An R package for standardizing scientific names of plants and animals,”*Plant Diversity*, vol\. 45, no\. 1, pp\. 1–5, Jan\. 2023\.[doi\.org/10\.1016/j\.pld\.2022\.09\.001](https://doi.org/10.1016/j.pld.2022.09.001)

\[14\] J\. H\. Caufield, C\. Kroll, S\. T\. O’Neil, J\. T\. Reese, M\. P\. Joachimiak, H\. Hegde, N\. L\. Harris, M\. Krishnamurthy, J\. A\. McLaughlin, D\. Smedley, M\. A\. Haendel, P\. N\. Robinson, and C\. J\. Mungall, “CurateGPT: A flexible language\-model assisted biocuration tool,” arXiv preprint arXiv:2411\.00046, Nov\. 2024\.[doi\.org/10\.48550/arXiv\.2411\.00046](https://doi.org/10.48550/arXiv.2411.00046)

\[15\] Z\. Ji, N\. Lee, R\. Frieske, T\. Yu, D\. Su, Y\. Xu, E\. Ishii, Y\. J\. Bang, A\. Madotto, and P\. Fung, “Survey of hallucination in natural language generation,”*ACM Computing Surveys*, vol\. 55, no\. 12, article 248, pp\. 1–38, Dec\. 2023\.[doi\.org/10\.1145/3571730](https://doi.org/10.1145/3571730)

\[16\] D\. Scheepens, J\. Millard, M\. Farrell, and T\. Newbold, “Large language models help facilitate the automated synthesis of information on potential pest controllers,”*Methods in Ecology and Evolution*, vol\. 15, no\. 7, pp\. 1261–1273, Jul\. 2024\.[doi\.org/10\.1111/2041\-210X\.14341](https://doi.org/10.1111/2041-210X.14341)

\[17\] X\. Mi, G\. Feng, Y\. Hu, J\. Zhang, L\. Chen, R\. T\. Corlett, A\. C\. Hughes, S\. Pimm, B\. Schmid, S\. Shi, J\.\-C\. Svenning, and K\. Ma, “The global significance of biodiversity science in China: an overview,”*National Science Review*, vol\. 8, no\. 7, nwab032, Jul\. 2021\.[doi\.org/10\.1093/nsr/nwab032](https://doi.org/10.1093/nsr/nwab032)

\[18\] C\. Lin, B\. Liu, M\. Zhao, K\. Ma, and L\. Ji, “Catalogue of life China: Towards an index of known species present in China,”*The Innovation Life*, vol\. 3, no\. 3, 100141, May 2025\.[doi\.org/10\.59717/j\.xinn\-life\.2025\.100141](https://doi.org/10.59717/j.xinn-life.2025.100141)

\[19\] C\. D\. Brickell, C\. Alexander, J\. C\. David, M\. H\. A\. Hoffman, A\. C\. Leslie, V\. Malécot, and X\. Jin \(eds\.\),*International Code of Nomenclature for Cultivated Plants*, 9th ed\. Scripta Horticulturae 18\. Leuven, Belgium: International Society for Horticultural Science \(ISHS\), 2016\. ISBN 978\-94\-6261\-116\-0\.

\[20\] A\. Hinsley, A\. C\. Hughes, J\. van Valkenburg, W\. Stark, T\. Q\. T\. Bui, R\. Cheung, J\. Hauck, P\. Kasoar, M\. Lee, A\. Lavorgna, B\. Phelps, R\. Williams, A\. Lopez Garcia, K\. F\. Smith, and D\. L\. Roberts, “Understanding the environmental and social risks from the international trade in ornamental plants,”*BioScience*, vol\. 75, no\. 3, pp\. 222–239, Mar\. 2025\.[doi\.org/10\.1093/biosci/biae124](https://doi.org/10.1093/biosci/biae124)