Over-salience, counted by scripts/salience_counts.py, 2026-09-13

$ python scripts/salience_counts.py --work <run> --evidence <work> \
      --database toxicology --column label
toxicology: 139 question(s) in both arms, 3 repetition(s) each
the column the evidence keeps naming: 'label'

1. runs that put 'label' in the SELECT LIST, per arm
   blob           74 of 417 runs   (182 name it anywhere in the statement)
   selection      97 of 417 runs   (205 name it anywhere in the statement)
   ratio        97 selection against 74 blob

2. blob right and selection wrong, split on whether the question's OWN sentence names 'label'
   sentence names it    12 of 61  broke  (20%, n=61)
   sentence does not     7 of 78  broke  (9%, n=78)
   two-sided Fisher exact p = 0.0839

3. the competing explanation: the value is spread across sentences
   of 19 broken question(s), 0 name a value that appears in no other sentence
   (if the blob's advantage were breadth, this would be most of them, not a few)

The hand count this replaces. While the session was running, the same
effect was counted by eye as 37 select-list mentions against 9, and a
break rate of 29% (n=38) against 5% (n=78). Neither survived being
computed by a program. The numbers above are the ones to use.
