Epic: federated tournament tuning loop (crowd-sourced per-language config refinement)

[coaxk]

North-star vision for the tournament (Judd, 2026-06-04). The tournament isn't a one-off local tool — it's the engine of a continual, federated refinement circle:

1. Users run tournaments locally (subgen config sweeps on their own audio).
2. Results feed global stats (telemetry.subarr.com / D1 gets a tournament-results channel).
3. Aggregate → curated best-config recommendations per language nobody could derive alone.
4. Fed back to the userbase as subarr's shipped defaults/recommendations.
5. More tournaments → more data → recommendations refine. Forever, even at plateau.

The per-language tuning knowledge becomes collectively owned + ever-improving — a moat no static table matches, and content no one else has published (the research confirmed the literature gap).

Hard dependency

Gated on the QE summit (#TBD-QE). Aggregating clip-noise winners would pollute the global stats — the loop is GIGO until the judge can pick a real per-language ACCURACY winner (ρ≥0.6). Until then ship base-camp value (risk flags + safe default + clip_agreement dial), not per-language adoption.

Build order (once QE lands)

1. QE summit → trustworthy winners.
2. Telemetry: tournament-results channel — privacy-preserving (no paths/titles; same allow-list discipline as the current payload).
3. Worker aggregation + /v1/stats/tournament → global curated per-language configs.
4. subarr pulls curated recommendations back, offers as opt-in defaults (local tournament can override).
5. Iterate as data accrues.

Suggested release: Backlog / v2 (epic). Refs the QE summit + telemetry infra (C:\Projects\subarr-telemetry).