Most gifted plants do not die of neglect.They die of a bad window.

Plant Pact takes the plant, the person, the city and the window, reads the real fourteen-day forecast and the climate normals for that city, and tells you the odds — including the week it is most likely to die. Then it asks you to come back on day ninety and record what actually happened.

Place a plant See the ledge Star on GitHub

No account, no API key, no third-party service to sign up for. Your pacts are stored against an anonymous HTTP-only session cookie.

What it is for

Before you hand it over

Model the exact spot — north window, 3 hours of light, watered every ten days — against the real forecast for your friend's city, and find out whether you are about to gift a coffin.

When you hand it over

Send a care card that names the one thing most likely to kill it, what to change, and the exact date the projection turns. It is a Markdown file or a printable page, not a screenshot.

Ninety days later

Come back and record what happened. Every outcome is sealed into a per-pact SHA-384 chain, so the record of your prediction cannot quietly be edited afterwards.

Ten seconds

01
Pick the plant
Twenty-nine houseplants with light, temperature and watering ranges, taxonomy resolved against the GBIF Backbone.
02
Name the spot
Choose the recipient's city and drag the window's light hours. The projection recomputes on every move, from live data.
03
Commit the pact
Persist it with its odds, its killer and its due date. The row and its seal are written in one transaction.
04
Take the card away
Download a Markdown care card or open the printable version, and come back on day ninety.

Start with a plant Or drive it from an agent

cities in the corpus

8,040

real-weather placements

0.956

held-out AUC

58%

base survival rate

The base survival rate is the share of randomly sampled placements — random species, random window, random routine, real weather — that the labelling procedure calls a survivor. It is the honest reference point for a gift: a well-placed plant should read clearly above it, and a bad one well below. All four figures come from the committed model card and can be regenerated with npm run train.

What this is not

Not a plant identifier. It resolves names against a taxonomy; it does not diagnose disease from a photograph.
Not a survival guarantee. The probability is a model output on a real corpus, and the corpus itself is only 58% survivors.
Not veterinary or horticultural advice. Care ranges are approximate published guidelines, and toxicity flags are conservative and general.