Lyrics Operations · Quality & Coverage
Accuracy Scoring Engine
Score every lyric line against the reference text. Route only the outliers to an editor.
scored, every run
Every line scored. Editors only see the lines that need a human decision.
Batch input
| Line | Reference | Candidate | Score | Deviation | Timing† | Editor |
|---|
Scope note. Line-level confidence comes from a live model call that compares each candidate line to its reference line. If a live model connection isn't reachable in the environment this file is running in, the engine falls back automatically to a local text-similarity heuristic and labels the run accordingly — the badge above the title and the status line both say which mode scored the batch. † Timing drift is a simulated placeholder for this prototype — a production build would consume forced-alignment timestamps from the ASR pipeline instead. Reference and candidate lines above are original demo text, not licensed catalog content.
Hosting this elsewhere. The live-model path calls api.anthropic.com directly, which only works inside Claude's own artifact preview, where the request is authenticated for you. A browser can't safely hold an API key, and Anthropic's API doesn't accept unauthenticated cross-origin calls, so opening this file locally or hosting it on a public page will always fall back to heuristic mode unless you add a small server-side proxy: stand up a one-route service (a Cloudflare Worker, Vercel function, or similar) that holds your API key as a secret, forwards the same JSON body to api.anthropic.com/v1/messages with an x-api-key and anthropic-version header, and returns the response with CORS enabled. Then change the API_ENDPOINT constant near the top of the script to your proxy's URL.