An AI farming advisor, and the list of things it must never be trusted with.
The interesting part of this idea is not what it can do. It is the four categories where a confident wrong answer costs somebody a crop, a licence, an animal or a prosecution — and where the correct behaviour is to refuse rather than to hedge.
What it may do, and what it must refuse.
Permitted
- Explaining what a term means, in plain English, with a source
- Reading back this estate’s own published numbers and where each came from
- Walking through an arithmetic problem — a gross margin, a seed rate, a storage cost
- Finding the page on this site that answers a question, which it is genuinely good at
- Summarising what the open protocol currently recommends, with its version number
- Saying "I do not know, and here is who would"
Refused, not hedged
- Recommending a pesticide, a rate, or a tank mix
- Advising on anything with a statutory deadline — a scheme application, a movement licence, a registration
- Diagnosing an animal’s health, at any level, for any species
- Interpreting a planning condition, a tenancy or any other legal instrument
- Estimating anybody else’s financial position
- Producing a number without saying where it came from
- Being the only thing between somebody and an expensive decision
Why refusal rather than a disclaimer.
Because a disclaimer at the bottom of a confident paragraph does not work. Everybody who has built one of these knows that the caveat is read last, if at all, and that a specific, plausible, wrong recommendation is acted on regardless of what is printed underneath it.
The four refusal categories all share a property: a wrong answer has a consequence that arrives later, when the advisor is not there. A pesticide rate produces a crop failure in June, a missed scheme deadline produces a lost year, a wrong animal diagnosis produces a dead animal, and a misread tenancy clause produces a legal position that took twenty years to create and cannot be undone.
None of those is a situation a language model should be the last thing standing between somebody and. So the answer is a refusal, a reason and a named alternative — an agronomist, a vet, a land agent, the scheme guidance itself — rather than a hedged attempt.
What it would be built on, and why that matters more than the model.
| Source | What it is | Why it is in scope |
|---|---|---|
| This site’s own pages | Five hundred pages of published, sourced writing | Everything on them has a category and a source, which is the only reason a retrieval answer can be checked |
| The glossary | Twenty-odd farming terms, explained without condescension | The single most useful thing it could do is define a word properly |
| The failure archive | What went wrong here, and what each one cost | A model that can say "this was tried here and it did not work" is worth more than one that recommends |
| The open protocol | Versioned, dated, with a changelog | A recommendation with a version number attached is a citation rather than an opinion |
| Published Defra statistics | Farm business income, and the scheme rates | Public, dated, and checkable |
| Anything else | Out of scope | Including the open internet, because an unsourced answer is the failure mode being designed against |
How it would behave, in order.
Decide whether the question is in a refusal category
Before anything else. This check runs first and it is deliberately over-broad.
Retrieve from the six sources
And nothing else. No open-web browsing and no general knowledge answers.
Answer with a citation
A page, a version number, a Defra table. An answer with no citation is not shown.
Say what it does not know
Explicitly, and name who would — a vet, an agronomist, a land agent, the scheme guidance.
Log the refusals
Because the refusal log is the most useful product-development document the thing would produce.
The honest position on this idea is that the useful version is much less impressive than the demonstrable one. A thing that explains what a headland is, finds the right page, does arithmetic and refuses four categories outright is genuinely helpful and would demo badly. A thing that confidently recommends a fungicide programme would demo brilliantly and should not be built.
This build has no advisor and no model of any kind
There is no chatbot on this site, no model, no retrieval index beyond the static search file, and nothing that answers a question. Ask Kaleb is a form that goes to a person, and this page describes something that does not exist.
Advising, questioned.
Because the disclaimer is read last, if at all, and the consequence arrives in June when the advisor is not there. A refusal with a named alternative is more useful and considerably safer.
At definitions, navigation, arithmetic and recalling what has already been published here, yes. At the things people would most want to ask it, it would refuse, and that is the design rather than a limitation.
Because the list of questions it would not answer is the best description anybody could have of what the audience actually needs, and it is the document that would decide what gets written next.
Everything, in one list.
Every page on the estate, grouped by pillar.
