The cure for AI slop is a 1986 aircraft manual/kit
experiment-results-openai.md
1 KB·Raw on GitHub ↗
Per-category detail for the gpt-5.5 run
STE experiment results — OpenAI (gpt-5.5)
6 prompts x 4 conditions. Violations per 100 words (lower = cleaner). em-dash = true dash chars only.
| Condition | words | violations | per 100w | vs baseline | em-dashes |
|---|---|---|---|---|---|
| baseline | 649 | 23 | 3.54 | — | 0 |
| banwords | 608 | 13 | 2.14 | -40% | 0 |
| orwell | 532 | 9 | 1.69 | -52% | 0 |
| ste | 454 | 8 | 1.76 | -50% | 0 |
Per-category (per 100 words)#
| category | baseline | banwords | orwell | ste |
|---|---|---|---|---|
| long_sentence(>20w) | 0.77 | 0.49 | 0.38 | 0.0 |
| semicolon | 0.0 | 0.0 | 0.0 | 0.0 |
| contraction | 0.15 | 0.0 | 0.0 | 0.0 |
| passive_voice | 1.39 | 0.99 | 0.75 | 0.66 |
| ing_main_verb | 0.46 | 0.16 | 0.19 | 0.0 |
| nominalization | 0.46 | 0.16 | 0.0 | 0.22 |
| phrasal_verb | 0.0 | 0.0 | 0.0 | 0.0 |
| banned_word | 0.0 | 0.0 | 0.0 | 0.0 |
| marketing_adjective | 0.0 | 0.0 | 0.0 | 0.0 |
| modal_hedge | 0.0 | 0.0 | 0.0 | 0.0 |
| long_paragraph(>6s) | 0.31 | 0.33 | 0.38 | 0.88 |
Headline: STE cut violations by 50% vs baseline (3.54 -> 1.76 per 100 words). Ban-words 40%, Orwell 52%.