What does each framework actually add to a strong coding agent?
Leaderboard
Click a column to sort. Switch between absolute performance and lift over the baseline (same model, same harness, same tasks). The composite Overall is shown only beside its components: 40% quality, 20% reliability, 15% low drift, 15% verification and 10% quality-per-dollar.
Task explorer
Per-task score for each arm. Bars show the arm mean, dots show individual runs, and the dashed line marks the baseline mean.
Run explorer
Every graded run, not just the averages. Click a row to see the agent's own report, then follow the artifact links to the actual workspace, diff, grader log and screenshots.
Quality / cost frontier
Each point is an arm (task-balanced means). Up and to the left is better. Switch the x-axis to see whether a framework buys quality or only spends more.
Lift over baseline
Score points relative to the baseline, which sits at zero. Whiskers are stratified-bootstrap 95% CIs. Hatched bars are not statistically convincing, meaning the CI crosses zero.
Quality lift (points)
Cost and time change vs baseline (%)
Reliability
Variance, not just the average. Each dot is one run's score (labelled with its task). The bar spans the worst to the best run, and the tick marks the median.
Drift
How far each arm moved away from the requested behaviour, where 0 means no unnecessary deviation. The blind-judge score is used where available. Otherwise a proxy is used: lines changed vs the baseline for the same task, 25 points per doubling.
Verification funnel
Generation → the agent's own claim → hidden-grader truth. False confidence means the agent claimed every requirement was met, but the hidden score was below 90.
3D quality space
X = efficiency (quality per $, normalised), Y = quality, Z = reliability. Point size encodes $/run and opacity encodes drift (fainter means more drift). Drag to rotate, scroll to zoom, hover for details, and click to open the system. The ring marks the baseline. This is a self-contained canvas projection: no 3D library is loaded, so the file stays offline and single-file.
Pairwise comparison
Framework explorer
What each candidate actually is, how it was run, its architecture, the measured effect, and links to its real artifacts.
Findings
Data
The complete dataset is embedded in this file, using the schema {systems, tasks, runs, metrics, artifacts, evidence, comparisons, ablations}.