Methodology

Days to maturity, Scoville heat, pod length, disease resistance: each is stored as a claim that carries the sources behind it, and the build refuses a record whose numbers cite nothing. Today 62 sources underwrite 310 resistance claims and 863 recorded measurements across 510 varieties. The Sources block at the foot of a page lists exactly what supports that page, and the evidence line above it says how many of those sources are independent references rather than commercial catalogs.

We publish a sentence about evidence, not a confidence percentage. A score would imply a statistical calibration this project does not have, and you could not check it. A sentence naming how many sources exist and how many are independent can be checked against the list printed directly beneath it.

A missing value renders as a dash. It is never filled with a plausible guess, an average, or a number carried over from a similar variety. 379 of 510 varieties carry no documented disease resistance, and that absence means nobody has recorded resistance data we could cite. It does not mean the variety is susceptible. Reading absence as a negative finding would invent a result, so no page here does it.

The same rule removed fields that existed only in principle. A hardiness-zone field sat in the schema unpopulated on every variety until it was deleted, because a declared-but-empty field becomes a fabrication the moment something renders it.

Stable descriptors of a variety, like fruit color, need no method. Performance figures observed under particular conditions do, and the build rejects any performance measurement that cannot name the method behind it. This is why a catalog's stated maturity and a replicated trial result will never appear here as though they were the same kind of fact.

Where sources disagree, the disagreement is kept. Several claims of one measurement may coexist rather than being averaged into a single tidy figure that no source actually supports.

This is now enforced rather than intended. 82 observations from 5 published university trials are recorded against 7 named measurement protocols, and the build refuses a trial result that cannot name the protocol behind it. Those results are stored on the trial that produced them and never written back onto a variety, so a yield or a cull rate can only ever be read next to the site, season and method that produced it. One variety in this database recorded a 66 percent cull rate in a spring trial and 2.8 percent in a fall trial the same year, in the same state. Averaging those would invent a third number describing neither.

Concentrated evidence is fragile evidence. The build computes how much of the database each source underwrites, so the risk is visible rather than assumed. The largest single source currently accounts for 20.7% of all 2343 claims (bakercreek-catalog), the top three for 45.5%, and independent references for 31.8%.

328 of 510 varieties carry no independent reference: they rest on seed vendors, on member-run collections, or on the organisation that released the variety. We would rather show you that number than let it sit unmeasured, and reducing it is ordinary work rather than a finished achievement.

Judging that requires knowing what a source actually is, so this database records a source's format and its authority as two separate things. A book is a binding, not a credential. An edited scholarly monograph and a garden paperback share a format and share nothing else, while a seed list maintained by a membership counts as real primary knowledge without being disinterested. Applying that split lowered our own headline figure, because it stopped counting two sources we lean on heavily as independent. The number got worse and more truthful at the same time, which is the trade this page exists to make.

Several distinct varieties are sold under one name, and one variety is sold under several. Names are therefore recorded as typed claims about a variety, never as the variety itself. When a label points at more than one record, search offers a choice instead of silently picking one, and varieties that people commonly confuse carry explicit links saying plainly that they are not the same seed.

The trial record is where this costs us something. Across those 5 trials 80 entries were grown and AMSeed publishes results for 36 of them. The rest are listed by name with no numbers, either because the variety is not in this database or because the report's label points at more than one variety we hold and attaching the result would mean guessing which seed was in the ground. A trial that grew a plant called simply "Brandywine" tells us less than it appears to, and saying so is more useful than picking one of five candidates.

The database can generate far more combinations than deserve a URL. A crop and trait listing exists only where enough varieties actually carry that trait; a location page requires real zone and frost data; comparisons are curated rather than machine-multiplied into tens of thousands of near-empty pages. Filtered views are shareable through the address bar while pointing search engines back at the single canonical listing.

It is not a claim of completeness. AMSeed is early, the crop coverage is narrow, and the gaps described above are real. It is a description of the rules the code enforces, which you can check against any page on this site. Where our data is thin, the site is built to say so.