Rigour, where
you can see
it.
We build software the way researchers run a study: hypothesis, method, evidence, replication. That discipline is visible in the products rather than kept in a document nobody opens.
Four things we won’t trade away.
These are constraints rather than aspirations. Each one has cost us a feature at some point.
Methods on the page
The formula, the threshold, the assumption: visible right where you’re using it, not buried in a help doc.
Grounded in real sources
Every AI output is tied back to something real: a teacher’s story, a coded transcript, or an open dataset.
Replication by default
Every analysis exports the data, parameters, and seed. Someone else should be able to run it tomorrow and get the same result.
Built with the field
Every product in the roster has a partner credited. None of it would exist without their practice and patience.
If a result cannot be reproduced by someone who was not in the room, it is not finished.
Where we use it, and where we don’t.
We use machine learning in most of the roster. We are specific about what it is allowed to decide.
Suggest, then defer
Models propose codes, themes and story beats. A person accepts, edits or rejects each one. The suggestion is never the final state.
Show confidence
Every classification carries its confidence score, visible in the interface and included in exports. Low-confidence output is marked, not hidden.
Keep data local
Session and reading state stays on the device by default. Cloud sync is opt-in and regional.
Score people
No product in the roster rates a learner, ranks a participant, or produces a performance judgement about an individual.
Generate ungrounded
No model writes content without a source to check it against. If there is no source, there is no generation.
Train on your work
Partner data, transcripts and reading history are never used to train models. That is a contractual commitment, not a setting.