v8 item weights recounted to 32,890 (generation was 1,000 light)

Queue: derive the active model before the liveness gates

Progress card: stall detector follows the phase; queue shows one real RUNNING

Owner-only columns: policy A/B scores, MCQ format, judged timestamp

Do not paint the progress card before the queue can verify it

Stop presenting a dead run as live progress

Make the manage panel follow the version, and never leave stale controls live

The corpus size in the header belongs to the version

Fix the board: VERSIONS read KS_34137 before its const was initialised

Give v8 the revised corpus's real item weights

A requeue must never revive a status that was set on purpose

Make the manage panel work across versions, and fix what it did wrong

Clear the previous version's rows when a version has none of its own

Add a benchmark-version picker: v7 keeps the published board, v8 starts empty

feat: live evaluation progress on the board

fix: manage panel stuck on 'Loading…' and visible to everyone

feat: private evaluations + maintainer manage panel (delete / re-run)

feat: owner-only partial runs (item cap + subtask picker)

feat: accept closed-source (API) models, restricted to the maintainer

fix: queue panel showed 12 of 17 models, in alphabetical not run order

fix: render rows that lack the v2-only dimensions instead of crashing

Header: categories count matches the 13 leaderboard dimensions

Adopt v2 corpus: 13 dimensions, weights sum to 34,137, pre-v2 rows exempt from the new keys

Merge Medicine + Biology into Knowledge & STEM (13 -> 11 dimensions)

Windows mouse-wheel horizontal scroll; remove per-user queue cap

Fix trackpad horizontal scroll: disable back/forward swipe-nav (overscroll-behavior-x:none)

Improve horizontal scrolling: freeze Rank+Model columns, styled scrollbar, edge cue

Move Submit to top bar; highlight best score per task column

Add gated model submission: FastAPI backend + HF OAuth sign-in + submit form

Replace Gradio leaderboard with a self-contained static app (stable, framework-free, custom UI)