What we claim, what we do not, and how you can check.

The claim

Animind estimates the likely behavioural and communicative state of an animal from simultaneously measured signals, expressed as a probability distribution over a fixed set of behaviour categories, with the evidence for each reading exposed.

What we do not claim

• We do not claim animals have a language with a lexicon that maps to human words. No output of this system is a sentence an animal said.
• We do not claim the categories are natural kinds. They are a working ethogram-style vocabulary, useful because it is consistent, not because it carves nature at the joints.
• We do not claim species-level accuracy figures. In this release the behavioural priors come from published literature, not from a model trained on labelled data at scale — so a headline accuracy number would be unearned.
• We do not claim to detect pain, illness or emotion in any clinical sense.

Where the priors come from

Each species carries a versioned adapter, and each adapter states its provenance. In this release every adapter is derived from published ethological and bioacoustic literature rather than from proprietary training data. That provenance string is visible in the admin model registry and travels with every interpretation as the species adapter version.

How the system can be wrong

Confusable categories

Several categories share surface features. A defensive warning and a territorial signal look similar from a single clip; play and dominance display look similar without the interaction structure. The strongest confusable is always surfaced as an alternative.

Individual variation

Adult cats direct most vocalisation at humans and develop idiosyncratic repertoires. Species-level priors are weak for them, which is why the per-animal baseline carries most of the weight once it exists.

Context you did not record

The most informative fact is often something outside the frame. The system can only weigh what it was told, and reports how complete the context was.

Observer effect

Recording changes behaviour. An animal oriented toward a phone is partly responding to the phone.

Selection bias

People record unusual moments. A behavioural history built from recordings is not a random sample of an animal’s day.

What each number rests on

Every figure this product shows is a claim about a living animal, and they do not all rest on the same thing. A signal-to-noise ratio is a measurement. A behaviour category is a ranking produced by rules nobody has ever tested. Showing both as a bare percentage asserted a precision the second one does not have, so the product no longer does it: a reading may only be as precise as its basis, and the weakest tier is shown as a position in a ranking rather than a percentage.

The ordering is deliberate and it is the part worth arguing with. A comparison against the animal in front of you outranks a finding established on other animals, even though the second has more literature behind it, because it is a claim about your animal rather than about animals in general.

How this compares with other products

This posture is not unusually cautious for the category. It is roughly where the credible end of the market already sits, and the comparison is the fastest way to see what is and is not being claimed here.

Activity and health collars

Whistle, Tractive, Fi and PetPace do not claim to interpret what an animal means. They claim to detect change in an individual against its own history — activity, sleep, scratching — and to prompt a veterinary conversation. All carry a standing not-a-medical-device statement. That is a narrower claim than this product makes, and it is the one they have the data for.

Per-animal vocalisation apps

MeowTalk puts the individual model at the centre: the universal model is the acknowledged weak one, and accuracy is framed as something that improves for your animal as the app learns that animal. The honest unit of a claim is one animal, which is why the baseline comparison ranks above the cross-species inference here.

Apps that publish a headline accuracy figure

Where a figure is published it is for a narrow, defined task — a small set of emotional states — not for open-ended interpretation. A number attached to a defined task is a different object from a number attached to a nineteen-way ranking, and the two should not be compared as though they were the same claim.

Button boards

FluentPet and similar are careful not to assert the animal has language, and describe what they support as learned association. This product takes the same line: no output here is a sentence an animal said.

Calibration

A stated confidence is a claim about the world and should be checkable. Every interpretation can be paired with a reported outcome, and the platform reports observed accuracy against stated confidence, bucketed. A well-calibrated system’s observed rate tracks its stated confidence. Buckets with fewer than about twenty confirmations are marked unreliable rather than plotted, because a handful of points invites over-reading.

The behaviour vocabulary

Twenty categories, fixed. Each carries a baseline urgency, the states it commonly co-occurs with, what to observe next to disambiguate it, and the categories it is most often confused with.

A reading can come back as any of nineteen of them. The last, insufficient evidence, is never ranked against the others — it is what the engine returns when the evidence gate refuses them all, which is why you will see both numbers used here.

Reproducibility

The interpretation engine is deterministic: the same fused features produce the same verdict. Predictions are immutable and versioned — a re-analysis creates a new version and supersedes the old one rather than overwriting it, so what the system said and when survives. Every interpretation records the model version, the species adapter version, the provider chain and the analysis latency.