This is a small CPU model for closed-set candidate lexical retrieval, not a free-form translator. It ranks source-backed candidate forms from the current Kuku Possum multi-variety evidence edition and abstains when no score clears the frozen threshold.

Every result retains variety, orthography, source IDs, source anchors, review status and rights metadata. The model never merges the three varieties.

The published evaluation is deliberately described as a training-overlapping closed census. It contains 1,230 unique (variety, English candidate gloss) query clusters and 1,881 candidate sense records. Top-1 retrieval was 1230/1230 and top-8 was 1230/1230. This measures reconstruction of the indexed candidate evidence, not unseen-word generalisation, sentence competence, fluency or community acceptance.

Publication relies on the recorded operator authorization kuku-possum-operator-publication-authorization-v1. The operator represented that the required permissions were held and directed publication; no external permission document was supplied or independently verified. Source documents themselves are not included.

All linguistic rows remain candidates. Outputs must be labelled experimental and checked by an appropriate qualified speaker or language reviewer before use. The model is not authoritative and must not be represented as faithful sentence translation.

Dictionary edition: kuku-possum-dictionary-roth-bulletin7-primary-corroboration-v0.7.0
Dictionary manifest SHA-256: 30b776bed3f5ac18342848b196d1ab6ffae7d468e35382a282dff5b27350d081
Grammar edition: kuku-possum-grammar-comparative-identity-boundary-v0.6.0
Grammar manifest SHA-256: fb6281e240e685694a0df0f98393b640251bbf6a62802eb4bf0b2c62da19e6e4