The system does six things, in this order. Each one solves a specific problem, and none of them depends on the previous one being perfect.
Where does the data come in?#
Through the field assistant's phone, inside the house. The form writes to the phone first and the row appears immediately, signal or no signal. When the connection comes back, whatever was captured uploads on its own, in order and without duplicates. Nobody has to remember to sync anything.
That is not a convenience. A poultry house is a metal shed full of equipment: signal fails in there most of the time, and a system that requires a connection in order to record ends up being recorded in a notebook and typed up on Friday. Data typed up on Friday is no longer useful for deciding on Tuesday.
How does the system know whether a figure is reliable?#
Every record comes in with a reliability mark and, when something does not add up, with the flag that says so: an impossible daily mortality, a weight that goes down, feed intake that does not match the age, a bird count that does not square with the losses, a record identical to yesterday's.
A doubtful figure is never rejected. Rejecting it teaches the operator to lie to the form until the form accepts it, and then the system holds data that is clean and false, which is the worst possible state. What the system does instead is record the doubt and widen the band of whatever it projects on top. Uncertainty in the input propagates rather than hiding.
The system also looks for the explanation before raising the alarm: if the same record carries 38 °C, that high mortality already has a cause and the flag weighs less. An engine that shouts on every flock gets switched off in week one.
How does it project a flock's weight?#
It starts from the published curve for that flock's genetics, rescales it against what the flock is actually weighing, and pulls it back toward the standard while the evidence is still thin. It works from a new customer's first flock, with no history.
There is no projection without a band. The system cannot return a bare number: it returns the value and the range it expects the value to fall in. The band widens with distance to the horizon and with the quality of the input, and on screen it carries the same visual weight as the number itself.
Every projection is stored with the day it points at. When that day arrives, the actual value is recorded beside it. That is what makes it possible to answer "how often have you been right?" with a table instead of an assertion.
Who decides what matters?#
Management does, once, in its own language. Four things get declared: what you are optimising for, what can never be violated, who wins when two things clash, and how much the system may decide on its own.
The three layers do not mix, and the order matters: what was declared non-negotiable cuts down the possible days, the priority rules decide who wins among the days that remain, and the preferences choose within what is already allowed. No target for kilos or margin can buy a violation of a constraint.
If no date satisfies everything that was declared, the system says so and recommends nothing. Quietly relaxing the least important constraint in order to produce an answer would mean violating something somebody declared non-negotiable, without telling them.
Before a policy can be activated it has to be simulated against the company's own closed flocks, and the simulation is tied to the exact version that was simulated. You cannot simulate one thing and activate another.
How does the decision reach the person who has to carry it out?#
As a WhatsApp message to a specific person, with the instruction, the reasoning in one line, and a confirmation that only the recipient can give.
There is no "sent" state. An unconfirmed order is an open task and keeps showing up as one. That the message left the server says nothing about whether anyone read it, and even less about whether anyone did it.
Orders about what has already been measured and orders about what is projected go into separate queues. They do not compete for anyone's attention.
What happens when something is not routine?#
It escalates, with the context already assembled. First to the person doing the work, then to the supervisor, then to management, with hours of waiting between each step.
What the escalation is chasing changes with the date: before the day of execution it chases getting the thing done; after that day there is no point chasing it, and what gets closed out is the impact that already happened.
The queue is sorted by money, not by date. And when there is no price loaded to estimate the impact of something, that something is not worth zero: it is worth "not estimated", goes to the bottom of the list but stays visible, and says which figure is missing. Putting zero would send a thirty-million order to the bottom because a price is missing.
There is an emergency stop per company, visible at the top of the screen. It halts what has not gone out and flags what already has, because a delivered WhatsApp cannot be un-sent. A stop with exceptions is not a stop.
Can you find out, months later, why it was decided that way?#
Yes, and that is the part that cannot be added afterwards.
Nothing is deleted and nothing is rewritten. A corrected record does not replace the previous one: a new one is added and both remain. Months later you can open a decision and see what data existed that day, what was projected, under which policy it was decided and what happened afterwards. Any alteration made from outside the system is detected.
What exactly does the artificial intelligence do?#
A language model writes the reasoning in one line. It chooses nothing.
The decision is already made by the time the model gets involved, and it gets involved with closed facts in front of it. If the text it returns contains a number that was not in those facts, the whole thing is discarded and the template wording goes out instead. An order is never left without its reasoning, and the reasoning never carries an invented figure.
The expensive mistake a language model makes here is not the argument: it is the number, which reads perfectly natural and nobody catches.
Everything else — projecting, filtering, prioritising, spotting the anomalous — is deterministic calculation. The same data under the same policy produces the same decision, every time. If it were not reproducible, it could not be audited.
What does the system decide, and what does a person still decide?#
The system decides what is repetitive and verifiable: which day suits each flock within what management allowed, who does what, and what deserves to escalate.
Still a person's: the health judgement, the relationship with the plant and the buyer, what to do about something that has never happened before, and how much of all this to delegate. That last one is not a menu option: it is the first thing that gets declared.