About ConservaNode
Every week, the United States government publishes hundreds of pages of dense legal notices about endangered animals, shrinking habitats, and wildlife protection laws. Almost nobody reads them—because they are buried in small-print bureaucracy and Latin species names. ConservaNode turns those public records into a visual, plain-English web that anyone can explore.
1. From Dense Government Notices to a Living Wildlife Map
When federal agencies protect a river species, update rules for a coastal animal, or adopt international trade protections for a threatened one, they publish an official notice in the U.S. Federal Register. ConservaNode reads those public-domain notices, matches Latin scientific names to everyday English animal names, and connects every species to the places where it lives and the laws that protect it.
2. What Is Inside This Early Prototype (v0.1.0)
ConservaNode is an early prototype built from 1 Federal Register notice records (1 unique government documents). Across those notices, our system identified 2 wildlife species & plants, 3 habitats & geographic places, and 5 conservation laws:
What You Will Find Here
Official Wildlife Protection Rules: Endangered and threatened species listings, special 4(d) conservation rules, habitat range descriptions, CITES international treaty updates, and agency responsibilities published in the U.S. Federal Register.
What Is Not Here Yet
Illicit-trade cases: The notices above are regulatory rules and protections. Illicit-trade cases reported in U.S. federal press releases are now being collected and appear on the Cases pages, but that data is preliminary: machine-extracted and not yet verified.
3. Three Ground Rules You Can Trust
Many AI websites summarize documents without showing their work—or worse, invent connections that aren't really there. ConservaNode is built around three strict guardrails:
1. Reports of Public Records, Never Accusations
Every entry on ConservaNode is a neutral summary of a public U.S. government document (a Federal Register notice or a DOJ press release), linked to its source. Records report what those documents say. Where a DOJ release says someone was charged, the page words it as an allegation and names no one.
2. No Named People
ConservaNode is designed to exclude the names of individual people, suspects, vessels, and private citizens. Personal names are blocked at the database level and checked by an automated privacy scrubber on every build. If one slips through, report it and we remove it.
3. Honest Labels on Every Connection
We never use fake percentage "confidence scores." Instead, every link is badged STATED (written directly in the same sentence) or INFERRED (both appear in the same notice), and we avoid linking unrelated wildlife through giant country hubs.
STATED) vs. shared-notice (INFERRED) links. Right: How our graph avoids lumping unrelated animals together through huge country hubs.
4. How Accurate Is It? (Human-Checked Numbers & Honest Limits)
We tested our extraction pipeline (powered by a small, cost-efficient Gemini Flash model and checked against the global GBIF Biodiversity Catalogue and U.S. Census Gazetteer) against human-reviewed benchmark documents. Preliminary data is published before the formal evaluation gate is passed, and is labelled as machine-extracted and not yet verified. Each record passes automated quality checks first (see Data Quality & Audit Governance). On the benchmark below, the categories we publish scored against a ≥90% precision bar on species and places:
| What We Extract | Development Test (40 Notices) | Independent Heldout Test (10 Notices) | Quality Gate |
|---|---|---|---|
| Animals & Plants (SPECIES) | 100.0% precision (85 correct, 0 false alarms, 1 missed) | 94.7% human-verified (18 of 19 confirmed right) | PASS (≥90.0%) |
| Habitats & Places (PLACE) | 97.1% precision (68 correct, 2 false alarms, 21 missed) | 100.0% human-verified (13 of 13 confirmed right) | PASS (≥90.0%) |
| Public Agencies (AGENCY) | 100.0% precision (50 correct, 0 false alarms) | 100.0% precision (10 correct, 0 false alarms) | PASS |
| Conservation Laws (STATUTE) | 84.1% precision (37 correct, 7 false alarms) | 100.0% precision (7 correct, 0 false alarms) | PASS |
Known Limitations in Seed Release v0.1.0
- How Much of Each Notice Was Read: In this v0.1.0 seed release, 0 of 1 notice records were extracted from their published summary (title and abstract), and 1 records were extracted from the first 3,500 characters of the full notice. If an animal or state isn't listed on a notice page, it may simply appear deeper in the full document. (Full-document reading is prepared for v0.2.)
- Shared-Notice Links (
INFERRED): About 86.7% of connections in the dataset mean two items were discussed in the same notice (INFERRED) rather than linked in a single sentence (STATED). - Small Benchmark Samples: Our accuracy numbers come from 40 development notices and 10 heldout notices (the 10-notice heldout split was tested once and permanently retired so we never "teach to the test").
- Omitted Categories: Experimental categories for trade commodities (
COMMODITY) and enforcement action types (ACTION_TYPE) were less reliable in testing and are left out of this release.
5. Disclaimer
ConservaNode summarises public U.S. government documents. Records report what those documents say; they are not accusations or findings by us. A charge is an allegation — everyone is presumed innocent unless proven guilty in court. Data is machine-extracted, may contain errors and is provided as is, without warranty. This is not legal advice. ConservaNode is an independent personal project, not affiliated with or endorsed by DOJ, USFWS, NOAA, the Federal Register or any agency; agency names identify sources only.
6. Open Public Data & Contact
Public Domain Dedication: All underlying Federal Register notices are public-domain works of the U.S. Federal Government, and our compiled ConservaNode dataset is released under the CC0 1.0 Universal Public Domain Dedication. You can also inspect our cryptographic audit trail on the Data Quality & Audit Governance page.
Questions, Corrections, or Feedback: Spot a common name that could be clearer or a fact that needs a second look? Email the project maintainer at [email protected].