Exp 2 — Audio judge axis-score distribution audit
N records analyzed: 49
- with text composite: 49
- with audio composite: 31
Composite-score summary
| text | audio | |
|---|---|---|
| n | 49 | 31 |
| mean | 8.76 | 7.53 |
| median | 8.66 | 7.60 |
| min | 7.87 | 3.00 |
| max | 9.74 | 9.50 |
Text judge (11 axes) — per-axis score distribution
| axis | n | mean | median | %≥7 | %≥9 | %=10 |
|---|---|---|---|---|---|---|
cultural_anchoring |
24 | 7.12 | 7.00 | 83 | 0 | 0 |
form_as_finite_generator |
49 | 8.86 | 9.00 | 100 | 76 | 14 |
genre_register_collision |
49 | 8.24 | 8.00 | 100 | 35 | 4 |
long_fuse_plotting |
49 | 9.31 | 9.00 | 100 | 90 | 41 |
politeness_as_masking |
49 | 9.35 | 9.00 | 100 | 90 | 45 |
re_listenability |
24 | 7.71 | 8.00 | 92 | 4 | 0 |
status_as_live_wire |
49 | 8.90 | 9.00 | 100 | 69 | 20 |
structural_reveal |
49 | 9.33 | 9.00 | 100 | 96 | 37 |
surprise_at_specificity |
24 | 7.67 | 8.00 | 100 | 0 | 0 |
tonal_duality |
49 | 9.47 | 10.00 | 100 | 88 | 59 |
voice_violations |
47 | 8.00 | 7.00 | 83 | 47 | 45 |
Audio judge (5 axes) — per-axis score distribution
| axis | n | mean | median | %≥7 | %≥9 | %=10 |
|---|---|---|---|---|---|---|
character_distinctness |
31 | 8.45 | 9.00 | 97 | 55 | 13 |
consistency_across_episode |
31 | 7.65 | 8.00 | 77 | 32 | 6 |
delivery_match |
31 | 6.71 | 7.00 | 65 | 10 | 0 |
pacing |
31 | 7.06 | 7.00 | 77 | 13 | 0 |
prosody_variation |
31 | 7.39 | 8.00 | 87 | 13 | 0 |
Ceiling saturation comparison
- Text judge: %≥9 = 63%, %=10 = 28% (n=462)
- Audio judge: %≥9 = 25%, %=10 = 4% (n=155)
Audio judge is 0.4× more likely to give ≥9 than text judge.
Reading: text judge has a citation cap that clamps unsupported axes to 6. Audio judge doesn't. If %≥9 audio is much higher than text, citations are doing real work and the audio judge needs the same constraint.
Audio score histogram (axis-level)
1: █ 4 3: · 1 4: · 2 5: █ 7 6: ████ 16 7: ████████ 35 8: █████████████ 52 9: ████████ 32 10: █ 6