OP, Thanks for the writeup, genuinely interesting to see someone trying to attack the popularity-bias problem head-on rather than just tuning it. A few questions before I try it:
What's actually happening under the hood when you say "no AI"? If it's not ML-based similarity, what is the matching logic — metadata tags, personnel/label/producer overlap, something else? "No AI" is doing a lot of work in the pitch and I can't tell if it means "no black-box model" or "no data-driven method at all."
On the Vault: you ran Aretha twice and got >90% different tracks. How is that a feature rather than noise? If the underlying selection criteria are stable, two runs against the same artist should converge on a similar pool, weighted or reshuffled. A 90% swing sounds more like the tracks are being pulled from a low-precision pool than that there's a coherent notion of "70 hidden Aretha tracks worth hearing." What's actually constraining the selection?
How was the depth-level distinction (Essential/Balanced/Adventurous) validated? Genre-adjacency algorithms are notoriously easy to make *feel* adventurous while just reflecting whatever metadata taxonomy you built on. Curious what the actual basis is — audio features, collaborator graphs, something else.
Also, about the "real person, real discography" claim. How are you filtering out the flood of AI-generated tracks and ghost-artist catalogs that Tidal and Qobuz are increasingly full of? If The Vault surfaces obscure tracks by design, it's pulling from exactly the part of the catalog where synthetic content is least policed. Do you have a verification step, or are you trusting the streaming platforms' own filtering (which is known to be weak)?
Finally, about the claim of "no promoted content." Sonic Oracle depends on continued API access to Tidal and Qobuz. If either platform changes terms, throttles, or decides discovery tools like this cut against their own recommendation/licensing economics, what's the contingency? Roon at least has direct commercial relationships with those services — do you?
Not trying to be difficult — I'd actually try this if the mechanism holds up, since Roon Radio's popularity leanings are a real limitation. Just want to understand what's driving the recommendations before I hand it my library.
