Evidence
Choose a category from the left.
Choose a category from the left.
Methodology
PitchBook DealID is the unit. We keep completed startup VC, angel, seed, and equity-like rounds; we exclude grants, debt, crowdfunding, accelerators, M&A, buyouts, IPOs, and secondaries.
Output: screened VC rounds, stages, amounts, HQ geography, verticals.PitchBook DealSynopsis can count for deal-level classification when it explicitly states what the new funds are for. It is not shown as a public article because it is proprietary.
Output: internal baseline evidence.Candidate URLs come from PitchBook CompanyNewsRelation, Brave Web search, and multilingual/local-language search. PitchBook news snapshots are unioned because articles are added over time.
Output: deduplicated article URL candidates.The extraction stack includes Trafilatura, Tavily, and Exa. Trafilatura provides local extraction; Tavily and Exa provide managed extraction, cross-checking, and recovery when article text is difficult to retrieve. Full text is stored privately for research and audit.
Output: article text, URL, domain, title, date, extraction metadata.Gemma 4 31B (google/gemma-4-31B-it) checks whether each article is about the same startup and same funding round, and whether it contains explicit use-of-funds language for that round.
Output: verified target-round articles and proof quotes.Gemma 4 31B (google/gemma-4-31B-it) assigns locked taxonomy labels. A label needs source text. We do not infer use of funds from company description, sector, stage, investor rationale, or market-size claims.
Output: product/R&D, GTM, ops/capex, geography, hiring, and other taxonomy labels.Bars show category incidence among rounds with a captured use-of-funds statement. A round can count in more than one category. Unspecified means no statement was captured.
Output: clean deal-level percentages for the left bars.The public atlas keeps short source-linked proof windows, not full articles and not PitchBook synopsis text. A dot appears under a category only if that local quote supports the category.
Output: source-linked snippets for inspection.BGE-M3 embeddings and curated semantic prototypes assign sub-categories. UMAP arranges snippets so related wording is easier to browse. Read nearby dots as related examples, not as a precise distance score. The map does not measure budget size, causality, or whether the categories are natural.
Output: interactive evidence browser with clickable source text and explicit map caveats.Known limits. Public media coverage is uneven by region, language, sector, company salience, and deal size. The public site does not expose raw PitchBook text, full article bodies, API keys, raw classifier rows, or raw PitchBook identifiers.