Your collars are already a social network observatory. Here's what that unlocks.

How the next generation of livestock wearables can read the herd's social network — and what I built to prove it.

Live demo: herdlink.atodev.xyz · Source: github.com/atodev/herdlink

Virtual-fencing collars are one of the genuinely transformative ideas in pastoral farming. A GPS collar that steers cows with audio cues replaces physical fencing, enables cell grazing at a precision no fence line can match, and puts a sensor on every animal in the herd, around the clock. Farmers who use them don't go back.

But there could be a gap between the data these systems collect and the intelligence they deliver — and it wouldn't be a hardware gap. HerdLink is a working demonstration — a full herd simulation, a trained detection model, and a live social network view, running in your browser — built to show what closing it looks like.

The opportunity: the collar collects more than the alerts use

Every collar on the market sells health alerts, and most work the same way: monitor an individual cow's activity and rumination against a baseline, alarm on deviation. It's a solid, proven feature. But it leaves the richest signal in the data untouched — and the threshold approach carries a false-alarm problem that gets worse the more animals you watch.

Weather shifts every cow's behaviour at once. A storm halves grazing time; a heatwave elevates every temperature on the farm. Fixed per-animal thresholds either fire on all of it — and the farmer learns to ignore the app — or get detuned until they miss real cases. Alert fatigue is what kills monitoring products, and it is a direct consequence of treating a herd as a collection of independent sensors.

The solution: read the herd, not just the cow

A collar network is not eighty independent sensors. It's a social network observatory. Every position report is also a proximity observation, and from proximity over time you get the herd's social structure for free: who grazes with whom, which animals form cliques, who sits at the centre of the network and who at the edge.

This matters because social withdrawal is one of the earliest and most reliable sickness behaviours in cattle — frequently visible before fever peaks or rumination fully collapses. It's been documented in the animal-science literature for decades. No sensor needs to be added to capture it; it's latent in data the collars already produce.

HerdLink implements this end-to-end, with the network analysis vocabulary translated for the farm office:

Each pair of cows within 15 metres accumulates tie strength — the fraction of recent time spent together.
Each cow gets a strength score (her total social embeddedness), a community (her grazing clique, found automatically), and a clustering coefficient (do her companions also associate with each other?).
Crucially, the detection model doesn't just ask "is she peripheral?" — some healthy cows simply are. It asks "is she becoming peripheral, relative to her own baseline?" The change in network position is the signal, not the position itself.

These social features feed a trained model alongside the behavioural ones, and every feature is computed relative to the herd and relative to the animal's own history — never against a fixed threshold. That one design decision makes the system weather-proof: in a simulated heatwave or snowstorm the whole herd's telemetry shifts together, the distributions move as one, and the false-alarm rate stays at zero.

The demo makes this falsifiable rather than rhetorical. You inject a hidden illness into a random cow — nothing on screen marks her — and watch the detector find her. Measured results on herds the model never trained on: zero false alerts across 24 simulated hours of healthy grazing plus forced heatwave and snow, and every injected condition (illness, lameness, oestrus) caught with the correct suspected cause within two to four simulated hours. Adding the social-withdrawal feature measurably accelerated illness detection — the network sees her leave her clique before her vitals settle the question.

The impact if implemented

Earlier treatment. Hours matter in mastitis and lameness outcomes; a signal that leads the vitals is treatment brought forward.
Alerts a farmer can trust. A system calibrated to near-zero false alarms — and that explains each alert in plain language ("moving 60% slower than the herd; left her usual companions six hours ago") — keeps its credibility.
Missed heats become rarer. Oestrus shows up as restlessness and changed social behaviour; each missed heat costs a farmer a cycle — roughly three weeks of milk or calving window.
A genuinely differentiated product surface. Per-cow activity alerts are table stakes across the industry. A live herd sociogram — communities, centrality, withdrawal trajectories — is information no incumbent surfaces, built entirely from data every collar already collects.
A platform for what's next. The same network opens oestrus detection via approach patterns, dominance-hierarchy inference, and community stability as a herd-level welfare indicator — each one a software release, not a hardware revision.

The point

Getting a rugged, low-power, reliable collar onto an animal and steering a herd with it is genuinely hard engineering — the kind that takes years and defines the category. This isn't a critique of that work; it builds on it. The collars are already in the paddock, already reporting position on every animal. The opportunity is to ask more of the data they're collecting.

That next step is a modelling decision, not a hardware one, and it was provable in software — a browser-based simulation with a real trained model and verified detection behaviour — which is exactly what the demo is. The documentation walks through every component, including a plain-language guide to the network analysis for readers new to SNA.

The hardware got us a sensor on every animal. Taking the herd's social life seriously once you have that data is the natural next chapter.

HerdLink is a personal demonstration concept and is not affiliated with Halter or any collar manufacturer. Simulation parameters are grounded in published cattle-behaviour research.