The idea

A direct offshoot of Basilect Engine, swapping the signal source. Instead of what artists say, this mines what curators do: NTS Radio tracklists, where a single human chose to put artist A and artist B in the same hour of music. Curated co-occurrence is structurally different from algorithmic co-occurrence (a million low-effort playlists averaged together) and orthogonal to genre tags or collaboration graphs. Two artists are connected here if they recur together across curators with different scenes, biases, and eras.

The data engineering

This one is mostly a crawling and canonicalization story:

• Crawl — NTS’s public API caps pagination well short of the full catalog, so the crawler uses the master sitemap as its frontier instead, reaching episodes the API literally cannot page to. Full corpus: 86,600 episodes across 1,709 shows, zero errors, crawled politely at bounded concurrency.
• Canonicalize — tracklists give free-text artist strings (“Rhythm & Sound, Paul St. Hilaire”, “Tyler, the Creator”). Resolving those to canonical NTS artist IDs takes a full-string-match-first strategy with a two-tier delimiter split and article-prefix fallback, so band names with internal commas and ampersands don’t get shredded. 80.5% resolution across 1.5M strings.
• Compute & rank — 11 million distinct artist pairs across 147k artists, ranked primarily by distinct curator count rather than raw plays, so one obsessive show can’t fake a connection.

The README doubles as field notes on the NTS API’s quirks — broken search endpoint, pagination ceilings, which sitemaps are trustworthy — because the next version (thesis-fit scoring, intent-density weighting) will build on the same corpus.