🏆 Demo leaderboard static snapshot
Leaderboard — one PRD, framework is the only variable
Per-trial telemetry
Runs across apps
The harness builds the same spec with every framework — and it now runs many specs. Each run is graded against its own frozen PRD and test plan; cross-PRD scores aren't pooled (different specs, different difficulty).
Beyond this snapshot — the full Eval Studio
This page is a read-only leaderboard. The local studio (bun run studio) adds the whole workflow:
Rank not just frameworks but harnesses — Claude Code, the Codex CLI, and a bare baseline — on one board. Candidates that don't support the chosen harness grey out with the reason.
Pick target, harness, frameworks, worker/grader models, sandbox provider, trials and concurrency — then launch a real run, a zero-spend dry run, or copy the CLI command.
Watch each run stream provisioning → installing → building → grading, with cost-so-far and a cancel that tears down the in-flight sandbox.
Per-candidate composite + dimensions, a step-by-step adherence comparison, excluded trials with reasons, and frozen provenance hashes.
Every PRD test-plan step with pass/partial/fail, weighted credit and cited evidence, plus the blind judge's five criteria with samples.
A read-only audit of what the agent built (file tree, cold-start contract, scrubbed blind copy) plus a one-click live demo — boot the built app in a sandbox and get a localhost URL to click through.
Replay the full build transcript with session jumps, an error navigator, and an outline trace — or watch it stream live (redacted) as the build runs.
Five sandbox providers, swappable worker/judge models, multiple harnesses, design-system adherence scoring, and bring-your-own PRD targets.
