# Kuku Yalanji Model-Version Lexicon and Elder Benchmark

The isolated-word result accepts every canonical dictionary headword attached to the same English gloss.
The elder corpus is reported separately and is not pooled with dictionary words.

## Isolated dictionary glosses

| Model | Rows | Exact accepted | Token contains accepted | Mean chrF++ | Mean CER | Empty |
|---|---:|---:|---:|---:|---:|---:|
| step2770 | 297 | 14.14% (42) | 14.14% | 28.06 | 0.807 | 0 |
| step3120 | 297 | 14.48% (43) | 14.48% | 28.83 | 0.811 | 0 |
| step4155_guarded | 297 | 16.16% (48) | 16.16% | 30.38 | 0.782 | 0 |

## Elder sentences

| Model | Rows | Exact | Corpus chrF++ | Mean sentence chrF++ | Empty |
|---|---:|---:|---:|---:|---:|
| step2770 | 43 | 0.00% (0) | 29.18 | 29.63 | 0 |
| step3120 | 43 | 0.00% (0) | 29.26 | 29.66 | 0 |
| step4155_guarded | 43 | 0.00% (0) | 29.43 | 29.89 | 0 |

## Interpretation limits

- Dictionary exact match is strict canonical-form retrieval, not semantic adequacy for free sentences.
- The dictionary informed earlier corpus work; this is a coverage probe, not an untouched test sample.
- The 43 elder rows are rights-cleared and excluded from training, but were observed in earlier evaluations.
- Pairwise 95% intervals in the JSON are paired row-level bootstrap intervals, not estimates over speakers, texts, or training seeds.
- Near misses require human linguistic review; character similarity is diagnostic and does not establish grammatical correctness.