# Alice MT analysis pack (18 langs · persona=poet · R1)

Public-domain Alice opening — complete parallel set + truncation ablation + quality metrics for lang.ugrad fanout training.

## Files
| File | Purpose |
|------|---------|
| `ANALYSIS.md` | Quality report |
| `alice_mt_bundle.json` | All texts + metrics |
| `pairs_complete.jsonl` | Train/eval complete refs |
| `pairs_truncated.jsonl` | Truncation ablation |
| `complete_parallel.md` | Human-readable 18-lang |
| `metrics.tsv` | Spreadsheet metrics |
| `METHOD_SEMANTIC_PPL.md` | Proxy BLEU + semantic bands + unigram PPL |
| `semantic_ppl_metrics.json` | Per-lang numbers |
| `NEURAL_PPL_COMET.md` | Neural next steps |
| `CALIBRATION_THRESHOLDS.md` | DA/QE gates |
| `correlation_metrics.py` | Spearman/Kendall/γ/Somers + bootstrap BCa + jackknife a + permutation |
| `PERMUTATION_JACKKNIFE.md` | Jackknife acceleration + permutation tests |
| `SPEARMAN_ALTERNATIVES_BOOTSTRAP.md` | Rank alternatives |
| `precision_sweep.py` / `eval_comet_ppl.py` | Offline eval helpers |

## Train policy
- Positive labels: `pairs_complete.jsonl` only (T3 band)
- Trunc rows: aux / partial loss only
- Density: up-weight JA/KO/ZH so Latin length does not dominate steps
- Live fanout: finish EN source before MT; Latin budget ≥ 1.15× len(EN)

## EN complete
> Alice was beginning to get very tired of sitting by her sister on the bank, and of having nothing to do: once or twice she had peeped into the book her sister was reading, but it had no pictures or conversations in it, 'and what is the use of a book,' thought Alice 'without pictures or conversations?'