Exp 1 — Within-seed run-to-run autorater variance
N records: 49 N distinct seeds: 17
Seeds with ≥3 runs: 4
Per-seed variance summary
| seed (first 50 chars) | n runs | mean text composite | sd text | sd audio |
|---|---|---|---|---|
Two undertakers prepare a wake. They've never met … |
12 | 8.57 | 0.23 | 0.90 |
A daughter and her father wallpaper a small room w… |
7 | 9.11 | 0.30 | 2.15 |
A 1950s vicar and his bell-ringer realise the chur… |
5 | 8.38 | 0.48 | 0.64 |
Two telephone operators on opposite ends of a swit… |
4 | 8.14 | 0.16 | 0.00 |
Mean within-seed sd (text composite): 0.291 Mean within-seed sd (audio composite): 0.921
Reading: a sd of 0.3 means the same seed run twice can differ by ~0.3 in either direction — so a 0.6-point combined score delta is what we need to call a real lift, not noise.
Per-axis within-seed std-dev (text judge)
| axis | mean sd | max sd | n seeds |
|---|---|---|---|
voice_violations |
1.56 | 3.11 | 4 |
genre_register_collision |
0.71 | 0.96 | 4 |
re_listenability |
0.65 | 0.96 | 3 |
long_fuse_plotting |
0.58 | 0.84 | 4 |
cultural_anchoring |
0.57 | 0.69 | 3 |
tonal_duality |
0.56 | 1.00 | 4 |
status_as_live_wire |
0.54 | 0.65 | 4 |
surprise_at_specificity |
0.50 | 0.55 | 3 |
form_as_finite_generator |
0.50 | 0.89 | 4 |
politeness_as_masking |
0.48 | 0.72 | 4 |
structural_reveal |
0.47 | 0.50 | 4 |