day 1-2 matcher and resolver - #2
Merged
Merged
Conversation
weighted duration/title/artist score (0.40/0.35/0.25), accept >= 0.75, ISRC match short-circuits to 1.0, duration delta > 15s hard-rejects. 49-case suite covering feat/remaster/live/remix/cover/karaoke/accents; 2 xfail document the same-duration-cover artist-floor gap.
…n tests aiolimiter paces musicbrainz to ~1 req/sec; pytest-asyncio (asyncio_mode=auto) runs the async source/resolver tests; integration marker gates the live-API test behind --run-integration.
provider-first resolve: deezer search (relevance-gated) -> isrc/duration enrich -> musicbrainz canonicalize (isrc-anchored, else nearest-duration in the title/artist cluster) -> verify_match. a real-data probe showed the literal mb-first top-hit flow false-rejects 4/5 seeds; selecting the mb recording that matches the provider track fixes it. offline tests + opt-in live integration test.
cover_mismatch() flags candidates carrying a cover/karaoke/tribute marker the seed lacks (compared per-field, so a legit "Cover Me" title isn't penalised); applied at the Deezer relevance gate and as a COVER_MISMATCH hard reject in score_match. catches the superset-named case ("<artist> Karaoke") that token_set_ratio scores ~1.0. found via adversarial review.
…mbid the isrc anchor is trusted only when the MB recording's title + duration corroborate it (a provider-supplied isrc no longer self-certifies at 1.0), and a recording must be credited to the query artist by MB artist mbid (rejects tributes/sound-alikes, accepts collaborations, order-independent). artist lookup is best-effort: a transient failure degrades to the title/duration path. found via adversarial review.
erick-ti
added a commit
that referenced
this pull request
May 30, 2026
* chore: anchor python lib/ gitignore rule so web/lib isn't ignored
The Python distribution-packaging block's bare lib/ (and lib64/) matched the JS app's web/lib dir at any depth, silently excluding web/lib/{scores,seeds,utils}.ts from version control. Anchor both to the repo root; no root-level lib/ exists, so nothing else changes.
* v1.1 phase 1: static next.js showcase frontend
Next.js 16 (App Router) + TypeScript + Tailwind v4 + hand-authored shadcn-style
primitives. output: "export" => a fully static site with no server runtime: it
renders only the committed web/public/seeds/*.json (DECISIONS.md 2026-05-29 —
static-precompute), never the live backend.
- seed gallery (8 genre heroes + 2 vibe-steered variants) and per-seed results view
- result cards: rank, Deezer track-PAGE link (invariant #2), source chips +
high-consensus, CLAP-reranked badge, per-row rationale, expandable score breakdown,
view-raw-JSON dialog
- four-axis score breakdown with correct, distinct per-axis semantics (raw cosine +
fixed band for audio/vibe, [0,1] within-batch for combined, unbounded RRF for cultural)
- static coverage funnel + system-transparency/provenance panel
- types/recommendation.ts mirrors api/schemas.py + the export-only coverage/meta keys;
every nullable field guarded
* roadmap: mark v1.1 phase 0 (data export) shipped
erick-ti
added a commit
that referenced
this pull request
May 30, 2026
…tion (#18) * web: add vercel analytics + speed insights Wire <Analytics /> and <SpeedInsights /> (both @vercel/*/next) into the root layout so the deployed showcase collects privacy-friendly page-view and Core-Web-Vitals data. Both no-op anywhere other than Vercel; compatible with the static export. * readme: live showcase section + screenshot Add a top-of-README section pointing at the live static showcase (doppel-music.vercel.app) with a home-page screenshot, what the gallery renders, and the invariant-#2 note that only derived output ships — no audio is persisted or served and the embedding pipeline is never invoked publicly.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Builds the Day 1-2 matcher:
verify_matchscores a provider track against a canonical recording (weighted duration/title/artist, ISRC short-circuit, duration hard-reject), andresolveties Deezer fetch and MusicBrainz canonicalization to it — provider-first, since a real-data probe showed the literal MB-first flow false-rejects most seeds because MB's top recording is usually a different version than Deezer serves. Several rounds of review then hardened it against the matcher's core failure mode (silently embedding the wrong audio): cover/karaoke/tribute rejection, independent ISRC corroboration so a provider-supplied ISRC can't self-certify, and an artist-identity (MusicBrainz artist MBID) check that tells tributes apart from collaborations. Comes with ~110 offline tests plus an opt-in live integration test that resolves all five seeds end-to-end.