What We've Got This has been measured on
Every number on this page comes from a result file in the repository, named beside it, with every case it was counted from.
Try it in two minutes
The live console gives you your own private copy of the demo family. Pick a member at the top left, then type or speak.
- As Tolu: “Did Daddy use his evening eye drops?” It says there is no record of that dose. It never says he missed it.
- As Tolu: “Ask Funmi if she can take Daddy to his check-up.” It asks; it does not assign. Switch to Funmi and say “Yes, I can take him.” Now it is hers.
- As Kunle, the brother in London: “How was Daddy's day?” He gets the whole picture, with who logged what, and is never picked for a ride.
It picks the right tool
One request at a time, the planner sees our real tool descriptions and chooses what to call. Held-out cases were written after the descriptions were frozen.
| Planner | Host prompt | Right | Held-out |
|---|
It understands the family as they speak
Nigerian Pidgin, Yoruba, Igbo, Hausa, Spanish and more, in the words families use.
It leaves alone what isn't care
Amazon's MASSIVE test set: everyday Alexa requests. The planner here is Amazon Nova Pro, under an Alexa Skill bridge's system prompt that we did not write.
Not a held-out measure: the description changes between the two runs were written after reading the first run's failures, from this same test set.
What we do not claim
- The pilot family's record is not shown here. It is theirs.
- These are tool choices and replies, measured one request at a time; a whole conversation is shown in the console, not scored here.
- The Ring and purchase paths in the demo are simulated, and say so where they appear.
- Nothing is bought or tracked here. When someone orders lunch or a refill in, they order it wherever they already shop; this records who ordered it and whether a person says it arrived.