Why the part that decides is not a language model#
Deciding when to catch a flock moves money and happens every day. A system that makes that call with a language model cannot answer the only question that matters when the result goes wrong: why did it decide that?
Here the rule runs the other way, and it is not negotiable. Projecting weight, discarding the days that violate a constraint declared by management and ranking the ones that remain is done by deterministic calculation engines. The same data under the same policy gives the same decision, every time. The language model comes in afterwards, and it comes in to write.
The reason is practical, not ideological. The expensive mistake a language model makes in a poultry operation is not the argument: it is the figure. A sentence saying 3,500 grams where the calculation said 3,400 reads perfectly natural, nobody catches it, and the crew catches the flock a day late. A weak argument gets debated on Monday; an invented number gets executed.
The two places where the language model does come in#
It writes the reasoning behind a decision already made
When the system recommends a catch day or issues a field order, the decision is already made and the model receives closed facts in front of it: the projected weight, the band, the constraint that ruled out the earlier days, the indicator that moves. Its only job is to turn that into one line the person carrying it out will understand.
And it carries a fence that is checked, not merely intended: if the text it returns contains a number that was not in those facts, the whole text is discarded and the template wording goes out instead. The comparison is by value, not by characters — 3,400, 3400 and 3.400 are the same number; 3,500 is not. If the model does not answer, if there is no connection, or if the text is discarded, the order still goes out with its reasoning written from the template and the reason for the rejection is recorded. An order never goes out without a why.
It answers questions inside the console
Inside the system there is an assistant that knows which module you are working in and answers about poultry husbandry and about how each screen is used. It covers what today gets resolved by calling the area manager: what an indicator means, what an alert is watching, where something gets recorded.
It does not query your data, and that is a decision rather than a gap. An assistant that reads the database and writes the figure back is the shortest path to an invented number that reads normal. Figures come from the screen, which computes them with the same formula as the report and the alert, so the three can never disagree.
What the model is not allowed to do, and who does it instead#
| The task | Who does it | Why not the model |
|---|---|---|
| Choosing a flock's catch day | The engine, under the policy management declared | It has to be reproducible to be auditable |
| Reading the crew's reply | A closed vocabulary: DONE or CANNOT, and anything unclear escalates to a person | A "done on Monday then" would close a task nobody did |
| Projecting weight, conversion and mortality | The performance objectives published by the breeder, scaled to the flock's measured deviation | Every projection has to come out with its confidence band |
| Spotting an impossible reading | Rules and statistics over the record | The same input has to raise the same flag every time |
| Deciding whether the system learned anything | A statistical test with two conditions at once | It is what stops the system "learning" from three flocks |
| Formulating the ration | Your nutritionist | Not a software task |
| Diagnosing a health problem | Your vet | Also not |
What you get to do once this is in your operation#
Receive the why and not only the what. Every recommendation and every order arrives with the reason in one line, in the language of whoever executes it. That is the difference between an instruction that gets followed and one that gets argued about.
Set how much the system decides on its own. Per decision category: suggest and wait, issue with a window to veto, or issue. Management declares the threshold and the system honours it to the minute, because acting at four hours on something promised at six is the fastest way to make sure nobody delegates anything again.
Ask it to learn from your operation — and to refuse when it cannot. Past a certain number of closed flocks the system tries to fit its projections to your heat, your density, your genetics and your start, and it adopts that fit only if it proves it wins. It is rejected more often than adopted, and when it is rejected the screen says so. Before that history exists it projects from the breeder's published curves, and it says that too.
Check whether the system's judgement beat the house's. The report compares flocks where the recommendation was followed against flocks where it was not, by genetics, season and farm, and it includes the table of how they stood before the decision: if the flocks that were followed were already doing better, the later difference is selection and not merit. Below five flocks per arm it refuses to conclude.
Rebuild any decision months later. What data existed, what was projected, what was decided, under which policy, and what actually happened. Nothing is deleted and nothing is rewritten, so the question is answered from the record and not from the memory of whoever was there that day.
What it needs from you#
Three things, and none of them is an extra form. Steady recording in the field, which is the raw material for all of the above and the reason the system captures with no signal inside the house. History: a statistical fit specific to your company requires closed flocks, not promises. And a declared policy, because an engine nobody told what to pursue and what never to violate has nothing to decide with.
What this is not#
It is not an assistant that decides: it recommends, and you set how much it is delegated. It does not learn from another company's data, because it never sees it. It does not replace your vet or your nutritionist. And it does not promise a saving or an index: what it promises is that every decision can be reproduced and defended in front of a partner, a bank or an authority.
If you want to see what all of this looks like over a complete chain, the demo session walks through one flock's decision end to end, and how it works explains the engine's layers one by one.