The longitudinal intelligence layer.
A technical overview of the infrastructure that converts fragmented companion-animal observation into persistent, replayable, interpretive continuity across the ecosystem.
The interoperability surface of a longitudinal inference infrastructure for the pet ecosystem.
Companion Health is a longitudinal inference infrastructure layer for the pet ecosystem.
It synthesises fragmented behavioural, clinical, nutritional, insurance, and research observations into persistent health intelligence over time.
Companion does not replace source systems. It creates interpretive continuity across them.
A nine-layer stack with one cross-cutting orchestration layer.
Each layer carries a single responsibility. Layer boundaries are the contract — what each accepts, what it produces, and what remains its own concern. The Signal Engine is the structural and conceptual centre.
Nine layers form a vertical compositional hierarchy. Telemetry enters at Layer 1, is normalised through Layer 2, validated at Layer 3, and synthesised into signals at Layer 4 — the Signal Engine.
A cross-cutting orchestration layer operates laterally between Layer 8 and Layer 9, composing primitives into consumer-specific Briefs without inventing parallel concepts.
Five principles govern every design decision.
{{ p.body }}
Synthesis is not detection. It is reconstruction.
The engine does not watch for thresholds. It establishes what is normal for a single subject over time, measures sustained deviation against a cohort-adjusted baseline, and reconstructs how a signal emerged. Six operations compose it.
Seven primitives compose every object.
The ontology is the vocabulary of the system — it governs the data model, the API surface, the storage schema, and the consumer interface. Outside it, no terminology is introduced.
The Signal object is the system's unit of currency. Its schema is intentionally narrow — every field load-bearing, every claim disclosed.
The API exposes the ontology. It does not invent parallel concepts.
Two endpoint families, one worldview: primitive endpoints expose the seven objects directly; Brief endpoints expose consumer-specific compositions. New signal classes are added as values, not new endpoints.
A Signal retrieval returns the primitive in its canonical shape — narrow, composable, fully disclosed. No diagnosis, no recommendation, no consumer framing.
Replay turns a predictive system into an evidentiary one.
A signal that cannot be reconstructed cannot be defended. Replay reconstructs the day-by-day emergence of every signal — forensic, immutable, reviewable. Three properties compose it.
The same signal becomes different intelligence in different consumer hands.
One underlying signal composes into many Briefs — each weighted for the consumer's interpretive frame. A single intelligence layer serves an unbounded number of consumers; Companion never competes with the products that consume it.
Nine services turn observation into intelligence.
This is where the intelligence lives. Not a model, not a chatbot — a sequence of nine deterministic services between a raw observation and a delivered interpretation, each with a single responsibility. They populate the nine-layer stack and feed the Signal Engine at its centre.
Fetch becomes more valuable as the ecosystem grows.
Every observation contributed sharpens the cohorts every other subject is interpreted against. The architecture is not a pipeline that ends at delivery — it is a loop that compounds. This is the structural basis of the Companion Intelligence Network.
Network learning closes the loop: each consumer interaction and each new contributor refines the cohorts, which sharpens every future signal and interpretation. Fetch is the exchange through which this accrues — the Canonical Intelligence Exchange — and the reason participation compounds in value rather than merely adding volume.
{{ d.body }}