Working now · no model required
Identify a seed
You have an unknown seed — in a shipment, on a boot, in a soil sample. Describe what you can see and the list narrows. No account, no upload, works on a phone with a hand lens.
Thirty-eight invasive and high-risk species are loaded. The funded work takes that to one hundred, adds calibrated photographs of known seed to compare against, and lets the same interface accept a picture instead of a description.
What can you see?
How big is it?
Longest dimension. Against a ruler or a coin edge is fine — this is the fastest way to cut the list down.
What shape is it?
Turn it over. Many seeds are flattened on one face where they pressed against a neighbour.
Does it carry anything?
Wings, hairs, hooks and fleshy attachments are how a seed travels, and they are the most diagnostic thing about it.
What is the surface like?
Tilt it to the light. Under a hand lens the difference between smooth and pitted is obvious.
Was it inside anything?
If you have the fruit as well as the seed, this is decisive. Skip it if you only have loose seed.
How do you think it arrived?
Optional, and a guess is fine. Where you found it often narrows things as much as the seed does.
Skip anything you cannot see. Unanswered questions are ignored rather than held against a species.
All 38 species
answer anything on the left to narrow this
What this key can and cannot do
It narrows an unknown seed against thirty-eight species that matter in Hawaiʻi, using characters anyone can see with a hand lens. That is what keys have always done, and it needs no model.
It also declines to answer. If no candidate clears the reliability threshold, or two are too close to separate, it returns unresolved and shows a shortlist instead of a name. A wrong name entered into a reporting system is worse than no name — it closes a case that should have stayed open.
What it cannot do yet is confirm the answer. Confirmation needs calibrated photographs of known seed to compare against, and measurements taken from that material rather than from a book. Those are the seed reference library — and once they exist, the same interface accepts a photograph instead of a description.
How the two halves fit together →Where these characters come from
Published floras and taxonomic treatments. Fruit type follows from the family and is not in dispute — a legume is a legume — and gross seed form and dispersal are equally well established. Enough to narrow thirty-eight candidates to two or three.
What is deliberately not here is anything this project measured. No dimensions from a calibrated photograph, no surface at magnification, no reference images. Every record says so. Blurring the two would make the tool feel more finished and be worth less.
Why the key comes before the model
A model that ranks photographs needs photographs of known seed to learn from, and those do not exist yet for these species. The key does not — so it works on day one, and it keeps working when the light is bad, the seed is damaged, or there is no signal at the dock.
When the reference library exists, the photograph path joins this one rather than replacing it. Two independent routes to the same answer are worth more than either alone, because when they agree you can act on it.