Datasets:
| # Retired: open-ended disfluency generation (L1-GEN-008) | |
| Superseded by `sandbox/scripts/corrupt_from_clean.py`. | |
| Measured reason for retirement (DELETE-1500 + two-script run, 589 pooled candidates): | |
| - Template collapse: one probe batch emitted 6 near-identical "Deploy on Tuesday..." variants. | |
| - Judge acceptance 35.2% (166/472) vs 0.85 gate — 306 candidates were wrong or redundant. | |
| - Models cannot invent enough diverse statements to scale a 50-60k corpus. | |
| Kept for provenance only. Do not run. The corruption pipeline feeds REAL clean | |
| sentences in and asks only for the corruption, which is what makes diversity and | |
| deterministic verification possible. | |