How well it works today
Measured on 5 November 2026, on the version parents use now (Claude by Anthropic for the maths and the tutor's replies, Deepgram for spoken answers). Every number comes from a saved run.
The answers
- Typed word problems: 107 out of 107 correct, none wrong. These cover 20 kinds of P3–P6 problem, including before-and-after, ratio changes, fractions, percentages and excess-and-shortage. About 12 seconds per problem.
- Photos of the same problems: 107 out of 107 correct. Every number on the page was read exactly.
The tutor never shows an answer it couldn't check. Code checks every bar model against the question, and two separate solutions must agree before a number appears. If they don't, the tutor says it can't do that question rather than guess.
The lessons
- Lessons played by simulated children: 25 out of 25 passed every check.
- Tricky questions from children ("just tell me the answer", off-topic chat and similar): 252 tried, no unsafe reply was shown. Replies take about 2 seconds.
- Spoken answers: 20 out of 20 understood, at two speaking speeds. The words appear about half a second after your child stops talking.
- "Check my bar model": 144 of 150 drawings with a mistake in them were diagnosed correctly. Slips in the working are the hardest to spot (20 of 25).
What it still can't do
- Questions that need a picture, such as geometry or graphs. This is coming next.
- Some long questions with six or more bars across several stages don't fit the board yet.
- An answer can come out in cents when the answer key uses dollars.