Rootstock

Traditional scientific field data collection is expensive and structurally inadequate. Personnel costs consume 50–80% of grant budgets. Existing biodiversity datasets cover less than 7% of Earth's surface at 5km resolution. 75% of citizen science projects produce zero peer-reviewed publications — because they collect data without a specific research question.

Meanwhile, 250,000+ personal weather stations (Weather Underground), 30,000+ air quality sensors (PurpleAir), and a growing market of water and soil probes are already deployed at massive scale. These are distributed mini-labs sitting idle. PurpleAir data has already been used in peer-reviewed wildfire research — the data quality is there. The coordination isn't.

No platform exists that lets researchers define multi-parameter campaigns and match them to citizen-owned IoT devices.

Researchers create data campaigns — structured requests specifying what sensor data they need, where, when, and to what quality standard. The campaign model inverts the typical citizen science flow: instead of "collect data and hope researchers use it," researchers define what they need first.

Scitizens — our term for citizen scientists — discover campaigns relevant to their location and equipment, enroll their IoT devices, and contribute readings. Devices submit data automatically after enrollment, fundamentally changing the retention equation.

Devices are untrusted. Every sensor reading crosses a trust boundary. The platform authenticates via mTLS with a self-operated certificate authority, authorizes via OPA against a device registry, and validates every reading against campaign-defined quality thresholds — schema, range, rate, geolocation, anomaly detection. Quarantined readings are never deleted.

Data is exported with full provenance metadata, quality metrics, and campaign context — meeting FAIR principles so the data is actually publishable. Contributor identity is separated from observation data. Raw contributor locations never appear in public-facing datasets.

Context windows are the bottleneck. LLMs effectively work at smaller scale, but as the application grows it buckles under its own weight. The LLMs get confused, find multiple patterns to follow, and ultimately fail to help your project grow.

Rootstock's architecture keeps every task small and self-contained. The LLM never needs to reason about the whole system.

Every choice you give the LLM is a risk to the stability of the project. Don't give it more choices than absolutely necessary.

Rootstock uses identical file names, identical structure, identical conventions across every module. The LLM copies the nearest neighbor — and it works every time.