Upload a conversation you had with an AI assistant. The Observatory reports how language production, timing, initiative, response, semantic contribution and repair were distributed between you and it — measured from the exchange itself.
Layer 1 runs on the structure of the exchange — token ratios, turn-taking, latency, speech acts. It never needs to know the subject matter.
Learner curves and per-participant collaboration curves are theoretically separate and stay separate. There is no blended score, by design.
A measure that the data cannot support is reported as null with a quality flag explaining why — never quietly rounded down to zero.
What these measures do not establish. They describe observable interaction patterns. They do not measure learning, mastery, engagement, trust, factual quality or causal influence, and they should not be read or reported as if they did. Every analysis the API returns carries this notice with it.