For Labels: Signal Before the Story

as of 2026-08-13 (fleet re-measure)

CORRECTED 2026-08-25: figures on this page were updated to the fleet re-measure of 2026-08-13, and claims that could not be reproduced were removed or restated at their measured grain. This page therefore no longer matches its upstream source document, which has not yet been corrected.

canonical source: hosaka-fm/seen docs/hosaka/external/for-labels.md

web version: /labels/

A&R runs on conviction under uncertainty, and being early is the entire trade. By the time a chart position or a press cycle confirms an artist, the opportunity has usually already re-priced. hosaka is built to surface the same artist earlier — while the evidence is still underground and hard to fake — by fusing every source that has actually seen the artist move onto a single identity primitive.

Leading over lagging

We split evidence into two temporal orders and gate on the first. Leading signals — DJ set inclusion (132,404 clusters), radio airplay (796,328), underground/ independent retail sell-through (116,989), and repost graphs (45,442) — move before an artist is legible to the mainstream. Lagging signals, press coverage volume (452,834 clusters) chief among them, describe what already happened; useful for confirmation, never for discovery. An artist enters hosaka's emergence cohort — 119,095 clusters as of today — only on leading evidence. Press alone doesn't buy admission.

The dossier: convergence is the conviction metric

Any single signal can be noise — a friendly booker, a one-off spin, a bulk retail buy. Real conviction comes from convergence: the same identity lighting up across independent signal dimensions at once. As of today, 13,352 clusters have converged across 2 independent dimensions, 1,733 across 3, 257 across 4, and 40 across 5-or-more — that last group is the deepest-conviction tier hosaka currently produces, artists confirmed by five or more unrelated ways of watching at the same time. Each dossier carries a break_odds figure calibrated to convergence depth, not one flat percentage across the board — the odds attached to a 2-dimension hit and a 5-dimension hit are never presented as the same bet.

Provable earliness

Every admission is logged bitemporally in a detection-time ledger — 3,161,405 rows and growing — recording not just what we know but since when we knew it. A watchlist isn't a claim of foresight taken on faith; it's a timestamped one you can audit after the fact, which is what lets us show, not just say, that a name was flagged before the lagging channels caught up.

Reach: genre and geography without borrowing an ID

Genre classification is cluster-native, not bridged through MusicBrainz or Discogs tags — 4,334,514 genre-tagged rows across 845,889 clusters. Geographic coverage spans 422 distinct RA-tracked areas by area id and 79 continuously monitored live radio stations, so underground reach isn't confined to a handful of scenes.

The identity spine — and why the unresolved tail matters

Every artist, label, and song resolves to one cluster_id, a content hash, not a lookup ID. That's what lets sixteen source kinds converge on one entity instead of leaving sixteen partial pictures. MusicBrainz and Discogs IDs attach to a cluster when they exist, but they are nullable annotations, never the key. Clusters with no external ID yet aren't a gap to fill later — they're where the newest, least-catalogued talent currently lives, and they sit in dossiers and watchlists on equal footing with fully-catalogued ones.

What a label gets

A watchlist gated on leading signal, not lagging noise. A dossier per artist showing which independent dimensions have converged and the odds that implies. A ledger proving when hosaka knew, so earliness is demonstrable, not asserted after the fact. What you don't get yet: a finished self-serve product — the underlying substrate is live and deep; the commercial layer above it (metering, billing, packaged access) is still catching up, so today's engagement is direct and hand-fitted to a label's workflow rather than a signup form.