Source-Grounded Business Summaries
Software Sushi

Available on AWS Marketplace · Version 0.1.0

Data and evaluation scope

Luna authored 60 fictional business records: 20 scenario families rendered in English, Spanish and Portuguese. Deterministic person/date/company/amount substitutions expanded them to 3,000 related variants, 1,000 per language. All variants passed structural source/gold checks. The complete v3 model diagnostic covers the 60 original seed requests, not all 3,000 variants.

These are AI-authored, development-exposed records. Translations and substitutions are correlated. They are not customer documents, blind independent cases, or native-speaker acceptance data. Independent human acceptance count is zero.

An AI label review added an omitted third-party venue inspection commitment in all three translations. Original labels are retained privately. Three ongoing incident-log records are ambiguous between a future task and a standing duty; they remain in the primary score and are reported separately. Both v2 and v3 diagnostics use the same corrected labels and identical request text.

The metric compares action segment/person/ISO-date metadata, not semantic task correctness or summary factuality. The latter remain unmeasured. See the evaluation table and remaining failures. Full engineering records, fixtures and raw diagnostics are retained privately; this public summary does not expose the private source repository or customer data.