# Pair-nomination gate: pre vs post (#888)

pre  margin: -0.0328 (falsifiable=True)
post margin: -0.0298 (falsifiable=True)
verdict: PASS -- post margin -0.0298 >= pre margin -0.0328

## Fixture revision, disclosed
Scored against the REVISED label set (see `_revision` in pair_labels.json). Two pairs
were removed from `unrelated` AFTER the first scoring, because both violate the
fixture's own stated criterion: each pairs a Source transcript of one session with a
concept derived from another session of the SAME recurring meeting series. Both pairs
of that shape were removed, not only the one that drove the original failing margin.

## What the verdict does and does not say
Both margins remain NEGATIVE: this signal does not separate related from unrelated
pairs, and did not before the change either. The gate asserts only that chunking did
not make it worse. Making it separate is not this change's job.

## Set delta (descriptive -- a changed set is the point of the fix)
|pre n post| = 12, lost = 0, gained = 0, jaccard = 1.0000

## Truncation witness
pre:  0 multi-chunk doc(s), max cos = 1.0000
post: 2 multi-chunk doc(s), max cos = 0.9183
