What Peabody measures
A few plain counts from our own records, and what we do not claim. The counts have not been taken yet, so none are shown.
What we have checked, one by one
Each is a check we ran ourselves. None is a study of children learning.
How long children stay in each game, before our rebuild
- How many
- 482 plays across 19 games, in the 21 days to 27 October 2026 (12 games that were later removed are left out)
- Who ran it
- Peabody's own play clock (a play counts after 8 seconds of active play); no one ran an experiment
- When
- 21 days to 27 Oct 2026
- What it shows
- Median play was 91 seconds for signed-in children and 49 seconds for visitors; 19% and 33% were gone within 30 seconds.
- What it does not show
- Whether anyone learned anything. It counts time in a game, not learning, and visitors are not necessarily children.
- Record
docs/measurements/first-minute-2026-10.md
Whether the AI that reviews our stories agrees with stronger AI reviewers
- How many
- 13 stories and 4 deliberately broken copies, 17 reviews from each of 4 AI models
- Who ran it
- Claude Code agents acting as the judges, set up by Peabody's builders; no person and no child read them
- When
- 1 Oct 2026
- What it shows
- Which AI reviewer agrees closely enough to be used in production (Sonnet 5 passed; Haiku 4.5 failed).
- What it does not show
- That children like or learn from the stories.
- Record
docs/measurements/story-panel-2026-10-01.md
Whether the reading-along feature can follow a voice on a device
- How many
- 2 runs on 1 laptop, with a computer voice reading a 38-word kindergarten passage
- Who ran it
- Aaron, on his own MacBook
- When
- 2 Nov 2026
- What it shows
- The pipeline works end to end on a fast laptop.
- What it does not show
- Anything about a child's voice, a real microphone or any other device; the record itself says it is not a result yet.
- Record
docs/measurements/read-to-peabody-2026-11-02.md
What we do not claim
- We do not claim Peabody raises test scores. We have not measured that, and these counts cannot show it.
- We do not count children who read without an account, and we do not follow anyone around the web: every number comes from our own tables, never an outside tracker.
- We do not round up, drop the bad weeks or compare to other products.
What we have not measured
- We have not tested with children whether Peabody improves reading, maths or any test score.
- We have not run a study in a classroom, and no outside evaluator has looked at Peabody.
- We have not shown that the 3-day, 2-week and 2-month spaced checks cause children to remember; the research for that exact schedule in children is thin (docs/WHY-IT-WORKS.md calls it weak).
- We have not measured how a child's voice is followed when reading aloud, on any device.
- We have not measured whether children who stop playing a game learned from it first.
- We do not yet have the counts of children and books: the script that reads them has not been run, so none are shown.
How the spaced checks work
When a child gets a skill right, Peabody asks about it again 3 days later, 2 weeks later and 2 months later. A skill is “done for good” only if all three go well; a miss sends it back to learning. That is the only sense in which we say something stuck.
More: privacy, how we use AI, and what we can show teachers, and what we can't.