Animind AI
Sign in

Research collaborations

We need people who will tell us we are wrong.

The species adapters in this release encode behavioural priors from published literature. Priors encoded by engineers reading papers need checking by people who have watched the animals.

Email the research team

What we are looking for

  • Species adapter review. Read the priors for a species you know and tell us what is wrong, missing or overstated. The provenance and limitations text is as important as the numbers.
  • Annotated data with proper consent. Recordings with expert labels, collected under an ethics approval, that can be used with a documented provenance chain and a revocable consent record.
  • Calibration studies. Settings where ground truth is available, so stated confidence can be checked against observed accuracy rather than asserted.
  • Negative results. Cases where the system is confidently wrong are more useful to us than cases where it is right.

What we offer

  • Research plan access for academic groups, including bulk upload, annotation tooling and dataset export.
  • API access and webhooks for pipeline integration.
  • Named provenance on any adapter your work contributes to.
  • Exports that carry model version, adapter version and coordinate precision, so they are citable.

Data governance

Media never enters a training dataset implicitly. Eligibility requires an account-level model-training consent and a separate per-item consent, and is recomputed from live consent records rather than cached — so a revocation takes effect everywhere immediately. Every dataset carries an immutable provenance record.

Open questions we would like help with

  1. How much of an individual baseline’s apparent value is real, and how much is regression to the mean?
  2. Where does a per-species confidence threshold belong? A single global threshold is convenient and almost certainly wrong.
  3. What is the right way to report an inferred channel — an elephant rumble reconstructed from harmonics — so that it is usable without being mistaken for a measurement?
  4. Does surfacing the competing interpretation improve user calibration, or does it just add noise? We believe the former and have not proven it.