For researchers
Model output you can disagree with, in a format you can publish from.
The platform’s predictions are a hypothesis to test, not a label to accept. Everything is built so that a reviewer can see exactly what the model claimed, what your annotators said, and where they disagreed.
What the work looks like
Set up the project and its ethics reference
Permit or ethics references are recorded on the project and printed on every export, so a dataset never circulates detached from its approval.
Define your own ethogram
Codes, definitions, mutually exclusive state groups and modifiers. Ethograms version rather than mutate, so annotations made against v1 stay interpretable after v2 changes a definition.
Annotate independently
The model prediction is shown and clearly marked as a prediction. Disagreements between annotators are detected automatically and queued for resolution rather than waiting for someone to notice.
Export with provenance
Every export carries the model version, the species adapter version, whether a row was simulated, and the precision of its coordinates. Nothing leaves as a bare label.
What you get
Projects, subjects, observation sessions and field IDs
Bulk upload with per-file quality reporting
Ethogram manager with versioning
Annotation workspace with disagreement detection
Cohen’s and Fleiss’ kappa, with sample-size honesty
Coordinates exported at their stored precision, with the fuzz radius attached
CSV and JSON exports carrying full provenance
API keys and webhooks for pipeline integration
Model categories are not ground truth
In this release the species adapters carry priors drawn from published literature, not weights learned from labelled data at scale. Treat every model category as a hypothesis requiring expert review before it enters an analysis.
Where consumer hardware cannot capture a species’ primary channel — elephant infrasound, cetacean ultrasound — the modality is reported as absent rather than estimated.