Add provenance README

README.md

@@ -0,0 +1,53 @@
+---
+license: cc-by-4.0
+language:
+  - en
+pretty_name: VoxCeleb2 speaker-id to celebrity-name mapping
+tags:
+  - voxceleb
+  - speaker-identification
+  - metadata
+---
+
+# VoxCeleb2 dev speaker names
+
+Maps each of the 5,994 VoxCeleb2 dev speaker ids (`id00012`...`id09272`) to a
+celebrity name with a confidence grade and the evidence behind it.
+
+## Provenance
+
+Two independent routes, cross-checked:
+
+- **YouTube titles** (`youtube-title` source): titles of the first two YouTube
+  videos per speaker, fetched via the YouTube oEmbed API, with names extracted
+  and graded per title; a reserved third video was fetched only to adjudicate
+  conflicts.
+- **VGGFace2 identity_meta** (`vggface2` source): the authoritative
+  VGGFace2 `identity_meta.csv` names joined to VoxCeleb2 ids via
+  `vox2_meta.csv` (VoxCeleb2 identities are a subset of VGGFace2).
+
+Route agreement over the 5,994 speakers: **agree = 4,691**, **conflict = 137**,
+**vgg_only = 1,166** (VGGFace2 name, titles gave no usable evidence),
+**yt_only = 0**, **unresolved = 0**. 69 of the 137 conflicts were manually
+adjudicated with third-video evidence (source `adjudication`); the remaining 68
+stay unnamed. A 40-speaker random spot-check found 37/40 fully consistent; the
+3 flagged speakers (id00783 interviewer-vs-guest risk, id03812 own-channel
+inference only, id07169 first-name-only title) are downgraded to `low`.
+
+## Records
+
+`voxceleb2_names_full.json`: `{"idXXXXX": {"name": str|null, "confidence":
+"high"|"medium"|"low", "sources": [...], "evidence": str}}`.
+
+- `high` — both routes agree, or adjudicated with strong third-video evidence
+  (4,720 speakers)
+- `medium` — VGGFace2-only, authoritative but title-unconfirmed (1,166)
+- `low` — adjudicated on weaker evidence or spot-check-flagged (26 named);
+  82 speakers remain `name: null` (68 unadjudicated conflicts, 14 adjudicated
+  to null)
+
+Coverage: **5,912 / 5,994 speakers named (98.6%)**.
+
+Consumed by [voxpath](https://github.com/DJRHails/voxpath) as the committed
+compact asset `src/voxpath/data/voxceleb2_names.json` (name + confidence only),
+attached to corpus per-speaker metadata.