Word-level evidence · counted, traceable, limited

Phrasebook · Concrete DNA · Machine-readable lexical index

25 disclosed patterns were searched across 66 recovered post captions and 47 original audio transcripts. There are 55 target-owned posts and 11 explicitly coauthored posts owned by another account. Counts below are the number of distinct post IDs containing a match in each channel. Caption counts include empty captions in the denominator; spoken counts use the 47 transcripts. They are not rates of the creator personally saying a word.

No translation, guessed synonyms or invisible lemmatization is used. Case is ignored; an asterisk means the specific prefix family documented in the JSON. A speech match must fit within one original Whisper segment. Repeated words inside one post still count as one post. Reels can reuse a conversation, so distinct posts are not necessarily independent episodes. The same source can appear in both columns.

Caption counts you can audit

The three “not hryvnias” captions and the three “randoms” captions are all target-owned. Six target-owned captions mention comments in giveaway/present contexts. This is a lower bound from recovered posts, not an all-account history. Their source lists are in the rows and JSON below.

Corpus concordance

Second-pass audio review

Nine source recordings were retranscribed with gpt-4o-transcribe-diarize, retaining the original Whisper transcripts unchanged. It provides local speaker labels and segment positions; those labels are not named identities and are never matched across recordings. The OpenAI transcription documentation describes the speaker-label output used here. The second-pass outputs stay in the private corpus; the publication includes conclusions and short examples, not complete transcripts.

What can be promoted into an authenticity brief

Caption-exact: the phrase occurs in recovered publishing metadata. Visible: the lettering or subtitle can be inspected in a linked screenshot. ASR agreement: two automatic passes retain a short line, still without a human listening guarantee. ASR only / disputed: use as a research lead or describe the function in paraphrase; do not sell it as a signature quotation. The strongest claims combine channels: the money euphemism has separate captions; the car accent has separate visual references; the numerical-question pattern has short turn-level speech examples. The weakest claim would be to flatten every transcript into one voice and rank its most frequent words as his catchphrases.

Reproducing the count

The repository script scripts/index_language.py reads the private corpus and emits lexicon.json. The export records the regexes, matched short forms, source IDs, caption/ASR channel, segment times and input hashes. It omits complete captions, complete transcripts, private filesystem paths and CDN URLs. The second ASR pass is intentionally excluded from the counts, so adding a verification transcript cannot double the reported frequencies.