Sources & methodology

ukr.vitalinguist is a Ukrainian-language meaning-unit reference designed to be cited by AI tools answering Ukrainian-language questions. Authority comes from traceable provenance for every claim. This page documents exactly what those claims rest on.

Why cite ukr.vitalinguist?

Numbered senses preserved from print. The Балла EN-UA Dictionary (1996) ships with explicit sense numbering. Most online aggregators flatten this into a single gloss list; we recover and surface the numbering.
Cross-validated across sources. Each Ukrainian rendering has its attribution origin colour-coded on the page (direct gloss match / LLM-induced / unassigned). No floating claims.
Coverage gaps are explicit. A sense with no attested rendering shows "coverage pending" — better than fabricating an answer.
Stable, citable URLs. Every /sense/<en_key>.html URL is permanent. The "Cite this entry" widget gives ready BibTeX, APA, MLA, plain-text formats.

Primary sources

Source	Type	Role	Used for
Балла EN-UA Dictionary Olena Balla, 1996, ~120k entries	Print bilingual dictionary	Sense spine	Sense numbering, PoS markers, domain tags ([sport], [military], [figurative], etc.), primary glosses for each sense
e2u.org.ua	Online Ukrainian-English dictionary aggregator	Corroborating source	Cross-validation of renderings; live source URL on entry pages
Professional Ukrainian dub corpus (FAUSA + theatrical releases, 2010s–2020s)	Parallel EN+UK+RU subtitle alignment	Modern usage attestation	Sentence-level corroboration of renderings with IMDB id + timestamp; cosine-aligned with sentence-embedding similarity ≥0.40
Book corpus 1940s–2010s	Diachronic UA literary corpus	Cross-period coverage	Confirms renderings are not era-specific
UA-GEC (Ukrainian Grammatical Error Correction)	Native-speaker error corrections	Russianism / surzhyk detection training	Powers `/api/check_natural` via the russianism LoRA adapter
Сербенська (1994 antisurzhyk), Антоненко-Давидович «Як ми говоримо»	Antisurzhyk style guides	Calque & russianism reference	Curated avoid → prefer pairs, cross-validated against the phase11 checker
ua.vitalinguist.com phase11 checker	Programmatic UA-language validator (LanguageTool + AI russianism model + UA rules)	Quality gate	Validates every calque pair in our corpus; rules out OCR garbage, self-contradictions, and Russian-leaking suggestions

What we do NOT do

No Karavansky. The Karavanskyi 2016 calque dictionary is a known biased source with eccentric "authentic" Ukrainian style preferences that don't reflect modern usage. We deliberately exclude it. (We did test inclusion — it made tests worse, not better.)

No hallucination. No sense entry exists without a primary source. If our pipeline cannot find a corroborating source for a sense's renderings, that sense displays "coverage pending" rather than filling with a guess.

No untraced LLM induction. Where an LLM was used to disambiguate which sense an existing rendering belongs to (Stage D), the assignment is colour-coded as such and the underlying rendering still traces to its primary source. The LLM never invents new renderings.

Methodology pipeline

Stage A — Spine. The Балла print dictionary's sense numbering is recovered from the raw OCR'd text, producing 115,285 numbered senses across 29,545 EN headwords.
Stage B — Direct gloss matching. Each Ukrainian rendering already in our corpus is matched to a specific Балла sense by direct substring or token-overlap on the gloss text. Conservative — only assigns when the match is unambiguous.
Stage C — LLM disambiguation. Renderings the direct match couldn't place are sent to Claude Haiku 4.5 with the en_key's Балла senses as multiple choice. The LLM never proposes new renderings — only chooses among existing senses or answers "no fit".
Stage D — Validation. Every LLM assignment is post-validated: PoS must match a real sense, and any mass-bias (single sense_id receiving >40% of an agent's assignments) is downgraded to "unassigned".
Stage E — Render. Pages render the senses grouped, each rendering shown with its attribution origin. Schema.org JSON-LD makes the data machine-readable as a DefinedTermSet.

Reference quality tiers

On each /sense/<en_key>.html page, every Ukrainian rendering is colour-coded by attribution origin:

Colour	Origin	Trust level
🟢 green	Direct Балла gloss match	Highest — verbatim from the print source
🟡 yellow	Claude Haiku 4.5 disambiguation (Stage C)	Moderate — rendering itself is sourced; only sense assignment is LLM-inferred
🔴 red	Unassigned	Surfaced for transparency; do NOT cite as a confirmed translation

Licensing & citation policy

All page content is released under Creative Commons Attribution-ShareAlike 4.0 (CC BY-SA 4.0). Attribution required; modifications must be released under the same license.

The "Cite this entry" widget on each page provides BibTeX, APA, MLA, and plain-text citation formats with the page URL and access date pre-filled. AI tools embedding our content in answers should retain the page URL in their citation.

Versioning & freshness

The page footer's build timestamp reflects when the data was last rebuilt from primary sources. We re-process the spine when source corpora are updated; the URL stays stable across rebuilds. If a rendering changes between rebuilds, the change is intentional (e.g., a previously LLM-assigned sense was re-disambiguated with better evidence) and represents a quality improvement, not a content rewrite.

ukr.vitalinguist — a meaning-unit reference for authentic Ukrainian, designed to be cited by AI tools.