Reliable LLM-assisted curation of ecological survey data: a deployed agent bridging field collection and data curation

What this was about

Andrew Tokmakoff described a deployed LLM curation agent for TERN Ecosystem Surveillance plot data. The data moved from an older system to a new EMSA-based, schema-driven app, and side-loaded data can bypass app validation. Declarative YAML rules compiled to SQL run first. The model then works through phased briefs and queries the database via a tool harness, surfacing ecologically implausible values (999 m hummock grass, shrubs taller than the canopy, inverted soil horizons). Findings go to a curator dashboard, and nothing is written without human approval. Lessons: adjudicate by class not row, never accept silence as 'no findings', and codify confirmed findings into deterministic rules.

Why it matters. It is a rare deployed, cautious example of LLM-assisted curation with clear governance (no autonomous writes) and a practical recipe others can reuse.

In the room

• TERN is Australia's government-funded ecological research infrastructure; the Adelaide team runs hundreds of revisited plots collecting soils, vegetation and samples.
• The new EMSA-based PWA is schema-driven, but a bridge is needed to the old publishing system; side-loaded Excel data lose app validation.
• Data can be schema-valid but implausible, e.g. a 999 m hummock grass placeholder. Such errors are rare human factors, not systemic.
• Declarative rules in YAML compile to SQL (e.g. point intercept heights > 80 m); rules were themselves created with the LLM.
• Phased prompts ('briefs'), e.g. flag a single species across 50+ intercepts and don't reason it away by bioregion (the model once dismissed 76 intercepts as normal for arid regions).
• Harness loop: brief and tools -> model requests queries -> harness runs SQL -> JSON rows back -> model reports findings -> findings table -> curator.
• Every database write passes through a curator; adjudication is by class, not individual rows.

Notable moments