How We Research

The four rules

There is no credentialed horticulturalist on this site, and inventing one would be the easiest lie in the category. What we do instead is make every number checkable. That comes down to four rules, and everything below is how they are enforced.

1. A value names its source. Not a bibliography at the bottom of the page — the individual field carries the publication it came from. Days to germination for tomato cites a specific extension page, and that page is one click away.
2. A value carries its confidence. “We read this in a university bulletin” and “this is a sensible default” are different claims, and they are labeled differently rather than blended into one confident voice.
3. Somebody else checks it. Whoever sources a batch does not sign it off. An independent pass re-opens the original sources and grades the work.
4. The gaps get published. A missing answer is reported as missing. It never gets filled with a plausible guess, and it never gets quietly dropped so the page looks complete.

What the audits found

This is the part most sites would not print. Every batch of plant data gets a checking pass before it merges — the sources fetched again and compared against what we wrote, by an automated checker that did not do the original sourcing. Across 260 batches, that pass has re-checked 12,315 values and recorded 191 problems — a finding rate of about 1.55%. Each one carries a written resolution in the batch record.

Being exact about that, because it is the sort of claim worth being exact about: 260 batches carry an audit record, and 126 of those name the independent checker that ran it. The rest record the count and the findings without naming who checked. These are automated checks re-fetching the cited pages, not a human proofreader — we would rather say so than let "independent review" do work it has not earned.

The shape of that table is the most useful thing we have learned about our own work. 160 of the 191 are miscites: the value itself was often defensible, and the page we pointed at simply did not say it. That is what careful sourcing fails at when you do a lot of it, and it is invisible from outside — nobody can catch it without re-opening every source.

The other 31 are the ones that matter, and some were genuinely wrong. A scuppernong grape had been written down as self-fertile when Clemson lists it as female-flowered and needing a pollinizer — plant that alone and it flowers and sets nothing, for years. A raspberry had a 5–10 gallon container against the cited Oregon State bulletin's 20–30. One plant height was marked source-verified on the strength of a quote that is not in the source at all.

The raspberry and the currant are both recorded as fixed before their batch merged. The grape's record does not say either way, so we are not going to tell you it never reached a page. That distinction is the sort of thing that gets rounded off in a sentence like this one, and rounding it off is how a page like this stops being worth reading.

We would rather show you the list than round it into the tidier number, because the tidier number is the one you should trust least. "All our mistakes were bookkeeping" is a claim to discount on sight. "We wrote down a grape that cannot fruit on its own, and the check caught it" is one you can actually weigh.

A finding is one recorded audit note, not one wrong value — a single note can cover several. One batch-211 finding covers nine candidate values rejected in one pass. Counts are generated from the batch ledgers, not typed by hand.

Every field says how well it is evidenced

A plant record is not one thing we either know or do not. It is dozens of separate claims at different strengths, and flattening them into a single confident voice is how gardening content usually goes wrong. So each field carries its own tier:

Counting every record we publish, crops and varieties together, 79.23% of values sit at that top "verified" tier — we opened the source and it states the value. Others below it still carry citations; that figure is only the strongest rung, not everything with a source behind it. Whatever a value's tier, it is labeled rather than disguised.

What we do not have

Every plant site implies complete coverage. Ours is not complete, and here is the shape of the hole. Of 39 core fields, only 17 are filled for all 152 crops. These are the thinnest:

The one at the top of that table matters more than the rest, so read it plainly: we hold human toxicity data for 7 of 152 crops, and pet toxicity for 66. Do not use this site to decide whether a plant is safe to eat, or safe around a child or an animal. A blank is a blank, not a clean bill of health. For that question go to your poison control center or your vet, who will give you an answer we cannot.

The others are ordinary backlog. Where a field is missing, the plant page says it is missing rather than printing something plausible.

What has to be true before a number publishes

Data is built in small batches — one crop or one field across many crops — and each batch runs the same path. Nothing skips a step because it is obviously fine.

1. Source it. Fetch the actual page, quote the actual sentence. Not a summary of a summary.
2. Record it with its provenance. Value, unit, source, tier — together, in one record.
3. Run the gates. 14 automated checks, described below. A block stops the batch.
4. Independent audit. Someone re-opens the sources and grades the batch. Findings go in the ledger whether they are flattering or not.
5. Ledger and verdict. The batch gets a written verdict, and the audit findings stay on the record permanently. The table above is that record, added up.

The 14 automated gates are the mechanical floor: they check the schema, the field shapes, the source registry, licenses, provenance, identifiers, cross-references, and the safety fence on anything health-related. A value whose source is missing, unlicensed or not in the registry is a block, and a block does not produce a warning — the build stops, deletes its own output and emits nothing. So there is no version of this dataset that exists having failed them; that is what "the mechanical floor" means here, rather than a score we passed.

What the gates cannot check is whether a cited page actually says what we claim. No schema check can read English. That is what the source re-fetching is for, and it is where every one of the 191 findings above came from.

What is in the dataset

Growth-stage rows are the unusual one. For each crop we record what the plant should be doing and when — germination, true leaves, flowering, harvest — with the signs you would actually see and the day it is late. That is the layer nobody else publishes, and it is what lets the app say your beans should have been up nine days ago.

Where the sources come from

A single “3,728 sources” headline would flatter us and mislead you, because those sources are not one kind of thing. Here is the honest split:

The full ledger — every source, its license, and what we are and are not allowed to do with it — is on the data sources page.

The deep dives

Some numbers need more than a paragraph. These are the full write-ups, method and measured error and all:

• How we compute frost dates — the 26-year Daymet surface, the elevation correction, and the national validation against 7,006 held-out NOAA stations, including the accuracy bar we set, missed, and had to change.
• Data sources and licenses — every dataset behind the site and the app, what it powers, and the terms it is used under.
• The frost-station extract — 7,305 stations of NOAA freeze normals, cleaned into one documented CSV and released free.
• Checks that pass on nothing — five times in two days our own automated checks passed while examining nothing at all, including one whose safeguard was defeated by the exact mistake it existed to catch. What broke, and why a green tick is only evidence if you have watched that check fail.

If we have got something wrong

We would rather hear it than not. There is no form to fill in and no account needed — email hello@gardentrack.app with the page and what is wrong, and it goes on the record the same way an audit finding does. If it changes a published number, the page changes and so does the note underneath it.

This is not politeness. On a site with no credentialed reviewer, a working correction loop is the actual quality mechanism — the audits catch what a second reader can catch, and you catch what only a gardener standing in the bed can.

The same applies to the parts we build in public. The build log carries the mistakes as they happened, including a frost warning that could never fire and three separate confidently-wrong answers about what you can plant in August.

FAQ