Milpa Gardens
Open the planner

For teachers and extension educators

You can point a class at this site because every claim on it carries an evidence grade, names the mechanism it works by, links the sources it rests on, and says plainly which of its own numbers it does not stand behind. Where it does not know, it says so instead of inventing a number.

Milpa Gardens is a garden-planning corpus - rules, folklore checked against the literature, and species records - written with an LLM and then held to a set of rules most gardening content, AI-written or not, is not held to. This page is those rules, in plain language, so you can judge whether it belongs in front of students.

What a grade means

Every enforced rule carries one of five grades. The word is the grade; the sentence is how far to trust it. These counts are read from the live corpus - 92 rules today - so this is the real distribution, not a claim about it.

Well established — 32 rules
Multiple independent, good-quality sources agree and the mechanism is understood. Safe to teach as established.
Promising — 47 rules
Good evidence and a clear mechanism, with fewer or less decisive sources. Reliable, not the last word.
A good hunch — 10 rules
A plausible mechanism with some support, short of proof. Useful if you flag it as tentative.
Contested — 2 rules
The literature disagrees, or the effect holds only under conditions the claim does not state. Present both sides.
Refuted — 1 rule
Widely repeated, checked against its sources, and it does not hold. Kept on the site so the correction is findable.

A grade is never given to a rule that cannot state its mechanism in one sentence without the word “helps”. A claim that only correlates, or works by a mechanism nobody can name, is downgraded or dropped — it does not get to sit at the top grade because it sounds right.

What “verified” and “unverified” mean

Verified means a human read a real source, named in the rule’s sources, and confirmed the claim — 90 of the 92 rules are verified today. Unverified means an LLM asserted it from general knowledge and no human has yet checked a source: it is labelled as such and never presented as fact. Right now 1 of them are unverified — the one rule that is not verified is marked as having no evidence at all, a human’s judgement that the literature is silent, not an LLM guess dressed up as fact. Promoting a claim from unverified to verified requires a person reading the source — the model is never allowed to promote its own work. That single rule is the difference between this and an AI that cites papers it never opened.

Numbers carry the same honesty at the field level: a species’ size, spacing or timing figure is marked estimated or contested when it is not from a firm source. Every one of those — 348 today — is listed on What we don’t know, the one page a gardening site is never supposed to publish.

What this corpus refuses to do

These are enforced, not aspirations — each is a line the project will not cross, for a reason a student can check.

How to cite it

Every rule, folklore and plant page carries a “Cite this page” block with a ready citation — and for a rule or belief its evidence grade is part of the citation. Please keep the grade when you quote it: cite a claim as what it is (“Milpa Gardens rates this Contested”), not as a bare fact, so your students see the confidence the same way you do.

What it is not

It is not a substitute for your local extension office or a soil test, and it is not safety or medical advice — a toxicity note is a general caution, not a poison-control reference. Its climate data (frost dates, hardiness) is from NOAA and covers the United States. It plans gardens from evidence; it does not replace an agronomist, and it says so.

You can also take the whole graded set with you: the downloadable corpus is every rule, its mechanism, its grade and its sources in one file, under a share-alike licence — something concrete to hand a class or check against yourself.

Read the rules · Download the corpus · Folklore, checked · What we don’t know