Working notes on what the published literaturedoes and does not establish.
Not a blog. These are the reading notes behind our product decisions, published because a company selling early warning should be checkable.
We write one of these when the literature changes what we are doing. There is no schedule and there will not be many. The first one exists because we had to correct a false claim on our own website, which is a reasonable indication of the standard we are trying to hold ourselves to.
August 2026
A pen-and-paper mortality log beats a bird-mounted sensor on warning time
Manual daily mortality records recovered one to five days of lead time on eight commercial broiler farms. Per-bird wearables give 14.9 hours. Electronic feeders with machine learning give one to three days. Hardware buys resolution, not horizon — with one important exception.
August 2026
Everyone says watch the water. Nobody has published how
Water is the most widely recommended early-warning indicator in broiler production. The only empirical test we could find produced 16.2 false alarms per 100 production days and concluded no reliable threshold could be established. The best available guidance is a duration rule, not a number.
August 2026
A broiler house has no normal. That is the hard part of the problem
Day of growout scores 20.391 in feature importance against 1.238 for temperature. Every expectation has to be age-conditioned, which is why threshold alarms cannot see slow drift. Three candidate baselines, and the false alarm budget that decides between them.
August 2026
Feed intake is the obvious signal. It is the one that failed in the house
Feed intake is the most intuitive early-warning variable and the one that failed hardest in commercial conditions. A commercial broiler preprint recorded 37.2 false alarms per 100 production days and no usable threshold. The 17-day lead from experimental pens is real, and it does not transfer to a bin.
August 2026
The sensors work in the papers. They are not in the house
Of 264 poultry precision livestock farming publications, 4 percent described a commercially available system. On 77 surveyed Tennessee farms, thermal cameras, microphones and activity monitors sat at zero while temperature, feed and water sat above 85 percent. Sound detection works in a commercial house. Distribution is the problem.
August 2026
A day earlier is worth something. It is not worth what people assume
USDA does not pay for birds that died from avian influenza, which makes the catastrophic case clean and rare. Lowering a mortality threshold fiftyfold bought two days in one model. Coccidiosis mortality was 0.9 percent of the UK cost. The payoff sits in the sub-clinical middle of a growout.
August 2026
It is not the temperature that kills. It is how fast you got there
A temperature-humidity index of 31.2 reached over six hours killed no hens. An index of 32.1 reached in one hour killed more than 95 percent in five. The water-per-degree coefficient remains unresolved across three published sources. A forecast plus a rate-of-rise rule beats a static threshold.
The two reference pages behind these notes.
Evidence base
Every figure we use, with its source, sample size, and age — including the gaps where no published figure exists. Read it
Competitive landscape
Who else is in this market, what they charge, and the funded attempts that failed. Read it
Research and pilots
The study we are trying to run and the partners we are looking for. Read it
Have a correction or a paper we should read?
We would rather be corrected than confident. Send it over.