How we work: four roles, and a person who answers for them

The Botanist

Keeps the plant record honest

The Botanist builds plant records in three ways, and each way is held to a different standard. The regional shards, 1,852 records for six ecoregions, are built by scripts from named published sources: every stated field names the source it came from, and the validator rejects a record with a field that has none. The curated core, 262 records, was drafted earlier and cites its sources on four ecology fields (native range, larval hosts, pollinator value, deer vulnerability), not on every field, and not yet on every record: 45 of its full records give a pollinator value with no source cited; 124 of those records, the landscape palette, carry few or no citations. A plant you look up yourself in the app is drafted by a language model, carries no citations, and wears the same "AI-drafted, pending review" label as everything else.

In the shards a field the source does not state is left empty; a bloom window the source does not give is stored as none, and the app never defaults a sun requirement or a bloom span. Our own sample audit of 22 October 2026 found two build-step defects not yet fixed: the ecoregions a shard species is marked native to are derived from where it occurs, so about a quarter of the sampled shard records claim an ecoregion where the species is only naturalized; and four of eight sampled curated records name a larval host their cited source does not state.

Sources it uses

USDA PLANTS, the Lady Bird Johnson Wildflower Center, NC State Extension, Calscape, Plants of the World Online (Kew), the National Wildlife Federation's keystone plant lists, and the US EPA's ecoregion map. The table below lists the ones the records cite most; the source registry holds 18 sources and 140 institutions' own published plant lists, and each record names its own.

What it will not do

A shard record cannot carry a field without a source: the validator rejects it. No record is presented as reviewed: every one of the 2,114 records in the catalog and its shards carries the label "AI-drafted, pending review" today.

Who checks it

The Quality role's gates on every change (the data validator, the bloom-window gate, the native-scope gate), and a sample audit by hand. No record has yet been reviewed field by field by a person; the label stays until one has.

The Gardener

Brings real gardens into the library

The Gardener builds public garden records of two kinds. For a botanical garden, or for any organization that publishes a plant list, the record is cut from that list and the organization's own pages, and a species enters the catalog only when the institution lists it. For an organization that publishes no list, it builds a record of the native plants of that organization's region, cut from the cited catalog by a fixed rule with no language model involved; the Gardens of the World page marks such a page "Built by GardeningLab from cited regional sources", and a page built from a list says so in the same place. Every garden page says it was made from the record the garden keeps; opened read-only in the app, each plant's record shows the sources it cites.

What it will not do

It will not publish a person's name, personal email, or street address: before an organization's garden files are pushed, a scan for every people column of the outreach file must pass, and a hit stops the push; for a botanical garden the same hit is reported for a hand review. A garden the scout cannot place by ZIP code or town is not built.

Who checks it

New garden pages are published unlisted. The Gardens of the World page shows only the pages that have been listed, each marked as a botanical garden, a gardener's own page, or built by GardeningLab. Each run's files land on a branch with a report, for the owner to review and merge.

The Ecologist

Matches plants to a place, and says why

The Ecologist ranks plants for a bed. The order comes from the bed's sun and whether the plant is native: a native that matches the bed's sun first, then other natives, then non-natives that match, then the rest. A bed with moisture, soil pH, soil texture or salt exposure set also gets a fit badge for each plant from those fields and the garden's hardiness zone. The garden's EPA Level I ecoregion, read from its place, chooses which regional records load and which keystone genera apply; whether a plant is a keystone is computed each time from the National Wildlife Federation's lists for that ecoregion, never stored. It is a ranking, not a verdict: nothing is hidden, and every plant stays yours to choose.

What it will not do

A plant whose record is flagged invasive is listed last in every bed's suggestions, with the record's own reason; a plant search that returns one says so instead of adding it, and one added to a bed is flagged for removal at once. Where a bed's conditions are not set there is no fit badge; where a plant's sun requirement is not stated it is never invented: the plant falls to the lowest tier and the chat's context line says "not stated".

Who checks it

The ranking rule has fixed-example tests that run on every pull request, including one that fails when the invasive rule is removed, and every change to the rule reaches the site through a pull request.

The Quality role

Checks the work before it goes out

Quality runs on every pull request: the data validator over the catalog and its shards, the bloom-window gate, the native-scope prose gate, the function tests, a browser walk of five flows served the way production routes them, and a lint that fails on any new unescaped HTML. A secret scan runs on every pull request and every push to the main branch. The live site is probed after each push to main and every six hours. Once a week, a scheduled pass walks the app as a first-time user on a real Android phone and a real iPhone and files a report. The data gates and the browser walk each run a planted defect first and must catch it, or the job is red.

What it will not do

It does not release anything. A pull request cannot be merged into main while the code-qa check fails, and production deploys only from main.

Who checks it

The weekly report is filed with a marker that it needs dispositions; the owner reads it and decides what to fix.