Methodology

Each finding starts as a question against the local OBIS/GBIF lake. The pipeline joins matching records with IUCN status, GEBCO bathymetry, and taxonomic concepts from the vault. An LLM or a human curator writes the synthesis.

The public surface ships only findings that pass the quality gate: no assistant residue, no placeholder paths, no ungrounded claims. The governed lake, raw OBIS dumps, and credentialed providers stay on the local Mac.

Severity

low / medium / high. The consequence of trusting a wrong answer — a misclassified threat label or population estimate. Flagship-conservation and quota-relevant findings sit at the high end.

Confidence (0–1)

Self-reported by the synthesis pass after grounding against the retrieved records. 0.8 = well-supported by sources; 0.3 = provisional, sparse corpus.

Evidence score (0–1)

Independent rubric over the cited vault references: source diversity, record count, recency, IUCN Red List entry. Computed after synthesis, not by the model.

n = (sample size)

Distinct knowledge-graph nodes cited. A loose proxy for depth — not for statistical power.

The hosted dataset is frozen at a dated snapshot. Live re-indexing runs on the local Mac; public release is cadenced. Cite the snapshot date.

Finding text: CC-BY 4.0. Records: per source — OBIS Core CC0, GBIF mixed per-dataset, GEBCO intergovernmental, IUCN Red List under attribution.

Not an official OBIS or GBIF product. A curated demo of a research workflow, not a regulatory dataset. No behavioural or telemetry data, no completeness guarantee in undersampled regions (deep tropical Indian Ocean, freshwater Asia, polar deep), no population-trend assertions without a cited primary source.