Knowledge tracing from tutoring transcripts

Predicting whether a pupil will answer the next question correctly, given the transcript of the tutoring session so far. Data from Third Space Learning and Eedi; evaluation by log loss, which makes calibration matter as much as ranking.

What makes this harder than standard knowledge tracing is that the signal is conversational rather than tabular: tutor and pupil turns with timestamps, uncertainty markers where the transcription was unclear, and background context that is not attributable to either speaker. Submissions run inside a container with no internet access, so everything has to be self-contained.