From a recording to an interpretation you can check.

The capture flow

1. 1

   Select or detect the animal

   Pick an existing profile, create one, choose a species, or record an unknown animal. Choosing a profile matters: it brings that individual’s baseline into the analysis.

2. 2

   Capture the observation

   Record audio, video or both, or upload existing files. During capture the app warns about low light, background noise, clipping, multiple animals and recordings that are too short to measure rhythm.

3. 3

   Add context

   What happened immediately before, during and after; whether food, a stranger, other animals or offspring were present; where you were and what the weather was doing. Context is not decoration — it is weighed.

4. 4

   Analyse

   The observation is queued and processed. Progress reflects real stage transitions written by the worker as each stage genuinely begins.

5. 5

   Read the result

   A primary interpretation with its confidence, the evidence for it, the evidence against it, the competing interpretation, what to watch next, and what the system could not measure.

The analysis stages

These are the stages the progress display shows. They are not a decorative animation: each one is written to the database when the worker starts it, and the interface shows an honest indeterminate state while a stage is running with no measurable sub-progress.

1. 01Detecting species
2. 02Isolating vocalisations
3. 03Analysing movement and posture
4. 04Evaluating environmental context
5. 05Comparing historical patterns
6. 06Estimating probable intent
7. 07Generating evidence-backed interpretation

What is actually measured

Pitch and its movement

Fundamental frequency across the call, its stability, and whether the contour rises or falls — bounded by the range the species actually produces, so an out-of-range harmonic is never reported as a fundamental.

Tonality versus noise

Whether the call is harmonically structured or broadband. This is the cleanest single distinction between, say, a whistle and a growl.

Rhythm and repetition

How many discrete calls there were, how long each lasted, and how regular the spacing was. A steady bark train is behaviourally different from an irregular burst.

Amplitude modulation

Purring, panting and trilling are amplitude-modulated at characteristic rates. The modulation is the signal, not the carrier.

Posture and movement

Body orientation, gaze direction, movement intensity and the body-language channels that matter for the species — a docked tail is a missing channel, not a neutral one.

Situation

The structured context you provide, mapped to a closed vocabulary the engine can weigh and cite as evidence.

Deviation from normal

How far this observation sits from this individual’s own running baseline, once enough observations exist for “normal” to mean anything.

How confidence is decided

A raw category score is never reported directly. It is reduced by every source of doubt, and each reduction is recorded so a result can be explained rather than merely displayed:

• Decisiveness — a 0.60 that beat a 0.58 is not a 0.60.
• Evidence volume — one signal is an anecdote.
• Contradiction — signals arguing against the reading cost more than corroborations earn.
• Data completeness — a thin observation can never look strong.
• Species certainty — an interpretation cannot be more certain than the species it rests on.
• Baseline maturity — a mature individual baseline earns a small, capped bonus.

What happens to your recording

• Media is uploaded straight to object storage through a short-lived signed URL; the API never proxies it.
• What you declared — size, type, checksum — is verified server-side before anything is analysed.
• Media never enters a training dataset implicitly. That requires both an account-level consent and a separate per-item consent, and revoking either takes effect everywhere immediately.
• Extracted features and interpretations survive media deletion, so your history stays readable.