A Registry-Driven LLM Pipeline for Evidence-Based Trait Extraction in Tropical Flora, Aquatic Life, and Exotic Pets

Abstract

This paper introduces a scalable pipeline that leverages large language models to generate structured, evidence-backed trait records for cultivated tropical plants, aquatic organisms, and exotic pets. The system ensures that LLM-generated data is auditable through four distinct mechanisms: a closed-vocabulary trait registry containing 39 versioned keys that restricts all input values to a typed schema; the inclusion of verbatim evidence quotes for each row to link data points directly to source text; a confidence rating system that assigns either "high" or "medium" status (excluding "low" confidence entries prior to storage); and the preservation of multiple versions.

When applied to 409,880 publishable species from the Tropical Species Encyclopedia, the pipeline completed 706,220 processing runs. It successfully persisted 5,489,881 trait records across 409,820 species, representing a coverage rate of 99.985%, with 81.57% of these records achieving high confidence.

We present three validation layers, ranked by evidentiary strength. At the full population level, 90.12% of the 5,427,588 rows containing evidence had quotes that were verbatim substrings of the source material (this figure rises to 93.49% if one compliance meta-trait is excluded). An audit involving a stratified sample of 100 non-red-zone rows showed that the quote supported the value in all 100 cases (lower bound: 96.30%). Additionally, a face-validity check on 50 red-zone rows resulted in 50/50 Accept ratings (lower bound: 92.86%). The study does not claim per-record correctness, noting that 100% of the data remains pending human curation. The primary contribution of this work is the four-mechanism framework.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC