Track annotation CSV

Ingest tracks.csv with one row per datapoint.

Columns: track_id, title, lyric_excerpt, proposed_genre.

Song JSON glob

Ingest songs/*.json — one JSON file per datapoint.

Map nested fields with dotted paths: metadata.writer, rubric.proposedMood.

Album cover join

Join annotations.csv to covers/* on album_id ↔ file basename.

Each datapoint shows the cover image plus CSV metadata (album, caption).

Liner notes split

Split notes/*.md on headings matching ^## TRK-\d+:.

Extract album metadata (title, release year) and per-track fields (writer, key) with regex.

Track rubric

Score each annotation on four dimensions with mixed scale types:

1. Validity — 7-point Likert (influence_gauge): does the genre tag fit?
2. Impact — ordinal: deep cut → iconic
3. Arrangement — ordinal: weak → definitive
4. Confidence — numeric 0–100 (certainty, advisory)

Weighted consensus

Genre fit consensus_weight = 2.0. Confidence weight = 0 (advisory only).

Conditional rubric

Skip Arrangement in the validator UI when Validity is False positive (consensus_skip.match; filler votes still count toward consensus).

Cover + liner note evidence

Validators see the album cover image first, then the markdown caption below with the album label.

Evidence order: image (cover path), then markdown (caption).

Lyric source link

Show source_link to the anthology at repository + edition, track sourcePath, line lineNumber.

Track metadata panel

Show song name, album, and writer in a datapoint_facts sidebar.

Genre label QC

label is ground truth — validators never see it.

Golden label field: label.

Prefill disagreement probes

Show label_shown as the model's proposed genre. Validators rate agreement.

Prefill hint: label_shown. Golden label: label (hidden).

Difficulty routing

difficulty (1–5) is metadata for analytics/routing. Strip from validator UI.

Difficulty field: difficulty.

Beatles catalog review

Add Junior and Senior critic classes with different reward_usd and priority.

Flagship albums (tier = iconic): larger panel with fixed senior share plus wildcard slots.

Tier-based routing

Match on proposed_tier: iconic → 5 validators (2 senior, 3 wildcard). Default → 3.

Wildcard composition

Route composition uses class = "*" so any qualified critic can fill remaining slots at claim time.

Senior count is fixed; wildcards fill the rest.

Escalation ladder

If the first round does not verify, add 2 senior critics, then another senior wave.

Ordered [[validators.routes.escalation]] steps — no per-step predicate.

Critic pay and stake

Junior critic: reward_usd = "5.00", stake_usd = "0.00". Senior critic: reward_usd = "20.00", stake_usd = "5.00".

Critic qualification

Require before claim: publication (required), musicology_credential (required), years_listening (optional).

Claim TTL

Critics have 30 minutes per claim ([validators.claims] duration_minutes = 30).

Conflict-of-interest decline

Allow critics to release an active claim without penalty when they reviewed the same album for publication.

[validators.actions] allow_conflict_of_interest_self_decline = true

AI critic panel

Run three AI classes per track (melody-model, lyrics-model, era-model) before humans.

Each class: type = "ai", model, and review prompt.

AI-first with human escalation

Initial route: 3 AI critics. Escalation: +1 senior human, then +2 senior humans if still unverified.

AI-safe scales

When AI critics are used, every rubric scale must use integer-friendly anchors:

• Likert: scale.size = 5 only
• Ordinal: 5 labels (or any count where (n-1) divides 100)
• Numeric: integer scale.values only