evals
Run evals on Jetty
v1 · May 2026 · jettyio · auto-evolved runbook (weakest-axis targeting)

opencode · Fusion router

opencode on Fusion has no image input, so it self-critiqued from SVG geometry instead of the render.

agent opencode · model openrouter/fusion

Judge, best round
28.3/40
round v2
Self-score, best round
36/40
inflation +7.7
Rounds
3
Jetty trajectories
opencode · Fusion router best round
best round v2 · bfa3612b ↗

Self-score vs judge, by axis

Best round v2, 0–10 per axis

agent self-scoreexternal judge

Judge: A clean, cohesive flat illustration. The pelican reads well from its long orange bill, throat pouch and S-curved neck, and the wing reaches the handlebar while the legs go to the pedals. Defects: the body is an egg-like blob with a detached tail wedge, the neck floats loosely off the body, the seat is missing or hidden, and the frame geometry is slightly odd, with a thin, oversized pouch outline.

Score by round

Out of 40. Each round starts from the previous round's best SVG.

self-scorejudge

This lineage had no image input, so its self-scores come from reading SVG geometry rather than the render.

Every round

Click a drawing to open it in the head-to-head viewer. The outlined card is the agent's best round.

opencode · Fusion router round v1
v1judge 29.3 · self 34 40df25cb ↗runbook
opencode · Fusion router round v2
v2 · bestjudge 28.3 · self 36 bfa3612b ↗runbook
opencode · Fusion router round v3
v3judge 28 · self 34 9776d557 ↗runbook

Scores per round

RoundPelicanBicycleCompositionPolishSelf totalJudge totalJudge spreadMinutesTrajectory
v17self 8.57.5self 87.7self 8.57.2self 93429.31–40df25cb ↗
v27self 96.8self 97.3self 97.2self 93628.31.5–bfa3612b ↗
v37self 96.7self 87.3self 87self 934281–9776d557 ↗

Want this kind of eval for your own agents?

Every run on this site is a Jetty runbook: versioned instructions, a pinned sandbox, and a trajectory you can inspect. Point Jetty at your task and get the same receipts.