Closed Measurement, Open Interpretation: A Four-Layer Behaviour-Classification Architecture and a Grounded Subconscious Reasoner for Companion-Animal Perception from Consumer Video
Year: 2026
Type: preprint
Abstract: Companion-animal products that speak in the animal's voice make factual claims about a living subject every few seconds. The characteristic failure of such systems is not measurement error but fabrication: an assertion that no admitted measurement supports. We present a production architecture that separates measurement from interpretation by construction — closed measurement, open interpretation — so that no path exists from absence of evidence to assertion. Every two seconds, a structured camera read is projected onto a fixed, version-controlled vocabulary of 19 channels and 216 classes with explicit abstention classes; out-of-vocabulary output is discarded at parse time. A 39-keypoint markerless pose engine and a facial-landmark channel supply measured channels as peers. Four layers classify: closed per-frame measurement; a salience-gated behaviour stream; 13 composite signatures and 18 named state transitions; and sequence models admitted to the user path only through a provenance gate that certifies trained status, on a declared dataset, for a declared species. Two principles organise the design: an admissibility principle — a claim may be asserted only if admitted evidence supports it, at sufficient confidence, recently, for the same animal in the same session — and a taxonomy of fabrication modes for percept-to-speech systems, each closed by a named architectural guard. A pain index mapped to the Glasgow Composite Measure Pain Scale — Short Form is bias-invariant by construction: in a demographic-invariance probe, varying demographics across 23 groups while holding behaviour fixed changed the score by at most 0.00. The measured channels include a gated kinematic engine whose false-locomotion rate on a seated animal is 1.8% of frames. A Reasoner, confined to a digest of sensor-derived facts and a strictly validated decision set, maintains the system's state record off the speech path; welfare and clinical language bypasses it entirely and is rendered deterministically. The first sensor-grounded utterance arrives 3–7 s after connection, and the transport layer sustains 100 concurrent sessions with zero failures. The validation is organised claim-first, in three evidence types: a demographic-invariance probe, a failure-mode taxonomy in which every mode carries its own guard check, and a latency and load characterisation of the production transport. The separation, and the guard catalogue, transfer to any embodied agent that speaks about a living subject.
Source: Figshare
Author: Yusuf Ghairat
Language: English
Subfield: Animal Science and Zoology
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7225465545