S.A.C.O. — the context filter

the problem

At compaction the agent keeps some segments of its transcript and drops the rest. Dropping a segment that mattered costs a wrong answer later; keeping everything costs the whole context.

Two numbers score a filter. F1 on keep measures how well it picks the segments worth keeping. R.95 is the share of tokens it saves while still keeping 95% of the segments that mattered — the number that decides whether the filter is worth running.

Every result uses the same held-out set: 13 sessions never seen in training. The first round scored candidates against the segments a summariser kept, and that label turned out to measure the summariser. The results page leads with the correction and the scores on an independent label.