Tools for the people deciding what to sign, book and play next — built on what a room actually did, not on what the numbers online say.
Know where a record sits against an artist’s or a label’s catalogue, in seconds — judged on how it sounds, not on who is posting about it.
Open FoundryThe moments of a set that actually landed, ranked by how the room responded — each one with the clip, so you can see it for yourself.
Open SignalWhere an artist or a record is heading, not just where it stands — built on what Foundry and Signal measure. Not built yet.
Coming soonSignal turns an hour and a half of video into a ranked read of the room: where the energy went, which moments it went to, and what was playing when it did.
A real report from the corpus above, not a mockup. Signal is a private demo — .
Ranked by what the room did. Not by streams, sales or posts — we have never seen those numbers, and nothing here is estimated from them.
The ten biggest measured responses in the corpus, ranked by crowd response during the moment. Track names are annotated afterwards by recognition and never feed the score — which is what makes it notable that 12 of the 24 named top-three moments are the headliner’s own record.
One set each, ranked by that set’s single biggest measured moment. This is a leaderboard of nine specific nights, not of careers — a different room on a different night moves the number.
Records that turned up in more than one set, matched on independently recognised names. Two, out of 150 identified segments — nine sets by nine artists in different scenes is close to a worst case for overlap.
Every set produces the data — the room ignites, stalls, and ignites again. nine71 turns that reaction into numbers that mean the same thing everywhere, so DJs, venues, and A&R teams can work from signal instead of vibes.
What we can show, what we cannot yet, and the measurements that told us which is which. Published including the ones that came back empty.
Moments are ranked by measured response; names are joined on time afterwards and never feed the score. So the two are independent, and agreement between them is corroboration rather than a circular check.
Suggestive, not validated. The label count is the reason.
Audio fingerprints should recognise the same unreleased record turning up in two different sets. The mechanism works and is switched on. At a threshold where precision is defensible it reports zero clusters on this corpus, and we publish that rather than lowering the bar.
Fixed by window coverage and a different corpus shape, not by a lower threshold.
Where a track starts and ends is currently decided by spectral change plus a minimum length — and a crossfade plays two records at once, which is exactly where spectral change detection fails. Recognition is the sharper boundary detector.
This, not better names, is the real argument for a paid recognition API.
The corpus, the charts and the validity ratios are generated from the reports by tools/build_charts.py, so they cannot drift from the runs they describe. The remaining figures are quoted from the runs named above.
We work with DJs, venues and labels who want a real read on the room — and on the records. If those are your calls to make, let's talk.