# Spine data dictionary

**Storage encoding (fix round 2, N2):** the `list`- and `dict`-typed columns below (`age_band_points`, `artists`, `composers`, `jury_rank`, `lyricists`, `source_urls`) are stored as JSON text in BOTH `spine.csv` and `spine.parquet`, not as parquet's own native list/struct types. Always read them with `schema.read_csv`/`schema.read_parquet`, never `pd.read_csv`/`pd.read_parquet` directly -- the `schema.*` readers decode the JSON text back into a Python list/dict; a bare `pd.read_parquet` call returns the raw JSON string for these columns instead.

| column | type | description |
|---|---|---|
| song_id | str | Stable slug: <san\|mel>-<year>-<title-slug>, numeric suffix on collision |
| contest | str | sanremo \| mello |
| year | int | Edition year |
| edition_number | int | Ordinal edition number |
| title | str | Song title as in the universe source |
| artist | str | Display artist string |
| artists | list | Individual performers |
| stage | str | sanremo: campioni \| nuove_proposte. mello: heat \| andra_chansen \| final |
| heat_number | int | Mello heat number, null otherwise |
| running_order | int | Start number in the stage |
| lyricists | list | Lyric credits |
| composers | list | Music credits |
| songwriters_raw | str | Verbatim credit string from the source |
| credit_split_known | bool | True only where the source marks text against music |
| reached_final | bool | Entry competed in the final |
| final_rank | int | Rank in the stage's final ranking, null if the entry did not reach it |
| stage_rank | int | Rank in the stage's final ranking, null if the entry did not reach it (equal to final_rank for a Sanremo row, since a Sanremo stage's final ranking is its only ranking) |
| outcome | str | winner \| finalist \| eliminated_heat \| eliminated_andra_chansen \| withdrawn |
| is_winner | bool | Won the edition |
| televote_count | int | Raw televote count where published |
| televote_points | int | Televote points where published |
| age_band_points | dict | Mello age-band points, JSON object |
| jury_rank | dict | Sanremo jury component ranks, JSON object keyed by component |
| televote_rank | int | Sanremo televote rank where published |
| vote_detail_available | bool | Any vote detail present |
| went_to_eurovision | bool | Represented the country at Eurovision. Null means unknown: no entrant-level source item exists for that country's edition to match against, not a known non-participation. False for a cancelled edition too (2020: a country's own selected entrant is known but the contest itself was never held, so nothing actually represented anyone). Review fix round 1, I6: the link can rest on ARTIST evidence alone for a national-final winner (no matching title candidate, but the same performer is confirmed to have represented the country that year) -- in that shape eurovision_rank/eurovision_year/wikidata_qid_song describe THAT Eurovision entry, which is not always the same song as this row's own title (e.g. an act winning the national final one year and being sent to Eurovision with a newly written song instead of their winning entry). A handful of confirmed cases of the latter are corrected to False via data/review/overrides.csv (reason=no_eurovision) rather than left silently wrong; a title-verified link does not have this caveat. |
| eurovision_year | int | Eurovision year. See went_to_eurovision's own caveat: for an artist-level link, this describes the artist's real Eurovision entry, which is not guaranteed to be this row's own song |
| eurovision_rank | int | Eurovision final placement. See went_to_eurovision's own caveat: for an artist-level link, this is the real entry's placement, not necessarily this row's own song's |
| wikidata_qid_edition | str | QID of the edition |
| wikidata_qid_artist | str | QID of the main performer. Always null in v1 -- population deferred, not yet built |
| wikidata_qid_song | str | QID of the song, usually null. See went_to_eurovision's own artist-level-link caveat |
| isrc | str | ISRC from Deezer or iTunes. Identifies a RECORDING of the song, not necessarily the exact performance from the contest -- a remaster, radio edit, live recording, instrumental, or later re-recording under the same title can all be the recording an isrc/preview_url points at (review fix round 1, I1/I2). Always cross-check preview_url/duration against the expected recording before using either for anything sensitive to the specific master (e.g. audio feature extraction) |
| deezer_id | int | Deezer track id |
| itunes_id | int | iTunes track id |
| preview_url | str | 30 s preview URL. Expires (hdnea token in the Deezer URL, an Apple CDN URL for iTunes) -- re-resolve via deezer_id/itunes_id before use, do not store or rely on the URL itself past the run that produced it. See isrc's own caveat: may be a different recording (remaster/edit/live/instrumental/re-recording) of the same song |
| streaming_match_confidence | str | high \| medium \| reviewed \| isrc, null if unmatched. high/medium: best_match's own algorithmic floors (title >= 85 always; high additionally needs both title and artist >= 90). reviewed: a human decision recorded in data/review/streaming_matches.csv, or the narrow safe-suffix-variant class auto-accepted from the near-miss band (title 60-85, artist near-certain) -- weaker title evidence than high/medium, confirmed correct per-row rather than by the algorithm's own strict floor (review fix round 1, Minor 5). isrc: deezer_id/preview_url were back-filled by smlc.enrich.spotify's cross-fill pass, which resolves Deezer's own ISRC endpoint for an isrc that came from Spotify (a Spotify search hit's own external_ids.isrc, for a row Deezer/iTunes never matched at all) -- an exact isrc match, not a fuzzy title/artist one, so treated as at least as strong as high |
| spotify_id | str | Spotify track id, resolved via the Web API client-credentials flow (smlc.enrich.spotify). ISRC-first: for a row that already has an isrc (from Deezer, see isrc above), resolved by an exact isrc:<code> search against Spotify's own catalogue -- Spotify's catalogue is not identical to Deezer's, so this can legitimately fail even for a row Deezer matched. Every other row (no isrc, or a failed isrc lookup) falls back to a title+artist search scored by the same streaming.best_match floors (title >= 85, artist >= 70) used for deezer_id/itunes_id. Carries the same isrc/preview_url caveat: identifies a RECORDING, not necessarily the exact contest performance |
| spotify_match_confidence | str | isrc \| high \| medium, null if unmatched. isrc: resolved by an exact isrc:<code> search, the strongest evidence available (not a fuzzy title/artist match at all). high/medium: streaming.best_match's own floors, exactly as documented under streaming_match_confidence -- a separate value space from that column, since a row can be isrc-confident on Spotify while only medium (or unmatched) on Deezer/iTunes, or vice versa |
| source_primary | str | Universe source for the row |
| source_urls | list | All source URLs that contributed |
| match_confidence | str | exact \| fuzzy_reviewed |
| notes | str | Free text, verbatim anomalies |

# Lyrics artifacts data dictionary

## lyrics_manifest.csv

| column | type | description |
|---|---|---|
| song_id | str | Spine song_id this row documents (1:1 with spine.parquet) |
| source | str | Lyrics source module that supplied this row: leparoledisanremo \| lyricstranslate \| letras \| eurovisionworld \| none (no source had a candidate) |
| url | str | Source URL the text was fetched from, empty for missing/ambiguous rows |
| retrieved_at | str | UTC ISO-8601 timestamp of the fetch, or of the most recent rebuild attempt for a still-missing/ambiguous row |
| sha256 | str | SHA-256 of the stored text file's bytes, empty unless status is found |
| source_language_label | str | Language label as stated by the source page. Often empty -- see caveats below |
| status | str | found \| missing \| ambiguous |
| char_count | int | Character count of the normalised stored text, 0 unless found |
| line_count | int | Line count of the normalised stored text, 0 unless found |

## lyrics_coverage.csv

| column | type | description |
|---|---|---|
| contest | str | sanremo \| mello |
| year | int | Edition year |
| stage | str | Same stage vocabulary as the spine (see the stage row above) |
| n_entries | int | Spine entries in this (contest, year, stage) group |
| found | int | Entries with a stored lyrics text |
| missing | int | Entries with no lyrics recovered |
| ambiguous | int | Entries where multiple candidates were found and no single one was confirmed correct |
| coverage | float | found / n_entries |

## lyrics_run_log.csv

| column | type | description |
|---|---|---|
| timestamp | str | UTC ISO-8601 timestamp of this build_all() run |
| found | int | Rows newly recorded status=found in this run |
| missing | int | Rows newly recorded status=missing in this run |
| ambiguous | int | Rows newly recorded status=ambiguous in this run |
| title_mismatch | int | Fetched pages this run whose own title failed verification against the spine title (see reconcile/store's _verified_title) |

## Known caveats

- 164 of 948 Mello `found` rows have an empty `source_language_label` (all 140 letras rows, which never populate this field, plus 24 lyricstranslate rows where the source page itself didn't declare a language). Downstream language-regime analysis should not rely on this field alone for those rows and will need its own language detection pass.
- 27 of 3145 stored texts contain a standalone bracketed section-label line left in the body (e.g. `[Refräng]`, `[Vers 1]`, `[Jojk]`, `(Chorus)`), mostly from lyricstranslate and leparoledisanremo. These will inflate token counts slightly if not filtered before use.
- Sanremo texts carry essentially no stanza structure (2109 of 2197 have no blank-line break at all) while Mello texts mostly do (5 of 948 lack one). Confirmed against the raw leparoledisanremo cache to be a genuine source-side property, not an extraction defect. Any stanza-level or line-block feature is reliably measurable on the Mello half of the corpus and not on the Sanremo half -- this asymmetry should be a documented caveat wherever such a feature is computed.

## Edition-level enrichment (added 2026-09)

| column | type | description |
|---|---|---|
| musical_director | str | Orchestra conductor / musical director of the edition (Sanremo "direttore d'orchestra / direzione musicale", Melodifestivalen "kapellmästare"). Edition-level, repeated on every entry of that edition. Null unless a single/principal edition-level director is named. Sourced from Italian (Sanremo) and Swedish (Melodifestivalen) Wikipedia edition articles. |
| md_status | str | Coverage flag for musical_director: `named` (a single/principal edition director is documented), `per_song` (each song had its own conductor, no single edition director), `no_orchestra` (playback, no live orchestra that edition), `not_stated` (the source does not report it). |