Reads: finalists; the persona and goal; signed-off claims and never-says (to spot conflicting past claims); competitor list; ../work-with-creators/references/vetting.md; ../work-with-creators/references/data-sources.md.
Does, per finalist — at the depth set in discovery (full, light, UGC rubric, B2B branch; see section 0 of the reference). Free first: everything below is done from the public profile and post pages in the host browser (reference §15 says what each platform shows logged-out), labelled "public-page read ". What a page hides (views on some platforms, comments behind a login, audience countries, transcripts) is listed for the paid top-up, never guessed.
- Read the profile and the last 12–20 posts on the main platform; drop pinned, giveaway and too-fresh posts; split sponsored and organic.
- Compute median views, view rate, engagement on the right basis for the platform, sponsored ÷ organic, mean ÷ median, cadence.
- Read the visible comments on 2–3 posts in the free pass (50–100 across ~5 posts in a top-up); record the share of generic/bot-like comments, self-descriptions, purchase questions, language and place signals.
- Audience fit: evidence and a confidence level (high / medium / low), from what people say about themselves — quoted and counted. Audience countries only for reach/install roles, top five at most. UGC-for-ads: skip audience; judge hooks (transcripts), product handling, past brand-ad work, portfolio, turnaround, honesty risk; mark delivery "not watched" and record 2–3 video links for the human.
- Brand safety: one web name search per shortlisted pick is part of the free pass — always run it (" scam / apology / drama / controversy"); then ~12 months of captions and a few transcripts; a name search with "scam / apology / drama / controversy" (if no search ran, the verdict says "captions and comments only"); past claims the brand could never make.
- Conflicts: competitor partnership ads, competitor tags in the last 3–6 months.
- Regulated and minors checks per the reference; any fail → out.
- Write the verdict line (the shape at the end of the reference). Fails become near-misses with the number that failed.
Writes: a verdict per finalist with all numbers, vetted_on date and post links; near-misses with reasons.
Rules: no number, no verdict. Label all authenticity findings "signals, not verified". Never infer ethnicity, age, gender, skin tone or health from images or names. No business-contact lookups here — they wait until the human approves the names.