A broiler house has no normal.That is the hard part of the problem.

Detecting an anomaly requires knowing what normal looks like. In most monitoring problems normal is roughly stable and the interesting question is deviation from it. A broiler house is not that problem. Over a growout of six or seven weeks, bodyweight rises by a factor of fifty, feed intake and water intake climb with it, target temperature falls on a programd curve, and the expected daily mortality rate itself changes shape — high in the first week, low in the middle, rising again at the end. Nothing is stationary. The baseline is a moving target that moves on purpose.

The evidence that age dominates everything

This is not a theoretical concern, and there is a clean empirical result behind it. A 2025 machine learning study in Poultry Science reported feature importances in which day of the growout scored 20.391 while temperature scored 1.238 and humidity 0.765.

Read that as a statement about baselines rather than about temperature. Almost all of the explanatory power is in knowing how old the flock is. Any signal you extract from a growout is a deviation from an age-conditioned expectation, not from a constant. A model that does not know the day number is not going to find anything.

Why the controller cannot do this

Eighty-four percent of US broiler houses had integrated electronic controls by 2011, and they are good at what they were built for. But a controller alarm is a threshold: this value is outside this band, act now. Thresholds work when the safe band is fixed and the failure is fast. Ventilation failure is exactly that case, and it is why the controller alarm is the right tool for it.

A threshold asks whether today’s number is acceptable. It cannot ask whether the last four days were acceptable given where the flock should be by now.

The conditions we are interested in do not present as a value leaving its band. They present as a value staying inside its band while drifting relative to where it should be for that day of that flock. Water consumption of a given volume is unremarkable on its own; the same volume on day 28 when day 27 was higher and the flock is growing is a different statement entirely. There is nothing in a threshold alarm that can express that, which is why every house can be fully instrumented and fully alarmed and still miss a slow drift.

Three candidate baselines, and their problems

If not a fixed threshold, then what. There are three practical options and none is clean.

The 4.3-house average comes from the 2011 USDA ARMS work, which remains the most recent national figure. We treat sibling-house comparison as the primary reference and the house’s own history as the fallback, and we say so because a grower has every right to ask what their numbers are being compared against.

The baseline moves between years too

One more complication that we think is underdiscussed. Between 2013 and 2025, US broiler mortality as published by the National Chicken Council rose from 3.7 percent to 5.95 percent, while feed conversion improved over the same period.

Whatever is driving that — longer big-bird programs, flock mix, disease pressure, all of which sit inside the number undecomposed — the practical consequence for a detection system is that a mortality baseline calibrated on a decade-old expectation is calibrated wrong. A threshold that was reasonable in 2013 is tight in 2025. Baselines have to be re-estimated from recent data, not set once from a reference table.

The false alarm budget

All baseline choices resolve into one number that matters more than accuracy: how often the system is wrong in the direction of crying wolf. The published work gives a clear ordering on that. On the same eight commercial farms, using the same method, daily mortality produced 2.0 false alarms per 100 production days, water 16.2, and feed 37.2. Only mortality yielded a threshold the authors considered reliable.

Against that, the dairy literature gives the tolerance: only 21 percent of automated alerts led a farmer to actually check the animal, and farmers act on fewer than roughly twenty alerts a day before disengaging. Those two findings together define the design envelope. A detection system for a broiler house is not optimizing for sensitivity. It is spending a very small false alarm budget as carefully as possible, and the choice of baseline is where most of that budget is won or lost.

What we do not know yet

Which of the three baselines actually performs best on a US broiler farm is an open question, and we are not going to pretend otherwise. The published lead-time work used a single fixed mortality threshold rather than comparing baseline strategies, and it did so retrospectively on Dutch farms during avian influenza episodes. Nobody has run this comparison prospectively on endemic conditions in US houses. That is a specific, answerable question, and it is one of the reasons our first commitment is a validation study rather than a product launch.