Reading the site

Leaderboard, frontier, heatmap, run detail

Every surface is built to be read at a glance and then drilled into. Here's what each one shows and how to interpret it.

Leaderboard columns

ColumnMeaning
#Rank among qualifiers; top 3 get gold/silver/bronze podium badges
ParticipantVerified GitHub identity (avatar + link)
PassedTasks solved on the ranked run
Billed costReal dollars from the ledger — the score
$/passedCost ÷ tasks solved — efficiency per solved task

Below it, a muted “Below the bar” table shows entries that ran but missed the eligibility bar, with a +N of how many more passes they need.

The cost-vs-pass frontier

Each entry is a dot: x = tasks passed, y = billed dollars. Read it like a shopping chart — you want down (cheap) and right (more passes).


   billed $
     ▲
  $11│        ◆ baseline
     │                       ● defunkt (many passes, costly)
   $8│              ● mona
     │        ○───────○ efficient frontier (dashed)
   $5│    ○
     │   ╱ ● pjhyett  ← "best value": cheapest that still qualifies
   $2│  ●   (down-and-right of the line = a better deal)
     │ ● (red = below the bar)
   $0└────────────────────────────────────▶ tasks passed
        55      59      63      67     71
MarkMeaning
◆ diamondThe baseline (GLM 5.2)
● green dotQualified entry (met the bar)
● red dotBelow the bar
○ dashed lineEfficient frontier — cheapest harness at each pass count

Analytics: the heatmap

Tasks (rows, hardest first) against submissions (columns, best first). Green = passed, red = failed, blank = not run. A red row across most columns is a task the whole field is stuck on; the hardest-tasks digest above it lists the lowest solve-rate tasks.

Run detail

Click any entry to open its run page, which hides nothing:

  • Pass/fail summary bar — solved fraction at a glance.
  • Per-task bar chart — cost per task, green/red by pass; click a bar for model + tokens.
  • Where the compute went — per-model tasks / tokens / cost.
  • The harness — the actual instruction + skill files, so others can learn from it.

See it in action from the leaderboard — click any row.