ops-dashboard · 15 runs · horizon 420s · 68 distinct screenshots · showing the first 210s, by when every run had ended
The chart stops at 210s because every run had ended by then. Each curve holds its last value from there to the 420s horizon, and the AUC is integrated over all of it.
Once the first build was on screen, every agent got the same follow-up prompts in its live session. Each edit is replayed from the moment its prompt was sent, so the panels show the same edit at the same moment however long each agent took over the first build.
| Run | AUC | Final | First render | First reviewable | By AUC | By final | Later changes | Window closed by |
|---|---|---|---|---|---|---|---|---|
| pi-sonnet-5.5 | 0.962 | 1.00 | 16.1s | 16.1s | 1 | 1 | none | signal |
| cc-sonnet-5.5 | 0.957 | 1.00 | 18.0s | 18.0s | 2 | 1 ≠ | none | signal |
| cc-sonnet-5.5 | 0.955 | 1.00 | 18.9s | 18.9s | 3 | 1 ≠ | none | signal |
| cc-sonnet-5.5 | 0.952 | 1.00 | 20.0s | 20.0s | 4 | 1 ≠ | none | signal |
| pi-sonnet-5.5 | 0.945 | 1.00 | 23.3s | 23.3s | 5 | 1 ≠ | none | signal |
| pi-sonnet-5.5 | 0.929 | 1.00 | 29.6s | 29.6s | 6 | 1 ≠ | none | signal |
| pi-gpt-5.5 | 0.929 | 1.00 | 30.0s | 30.0s | 7 | 1 ≠ | none | signal |
| pi-gpt-5.5 | 0.926 | 1.00 | 30.9s | 30.9s | 8 | 1 ≠ | none | signal |
| cc-opus-5.5 | 0.920 | 1.00 | 33.4s | 33.4s | 9 | 1 ≠ | none | signal |
| cc-opus-5.5 | 0.917 | 1.00 | 34.7s | 34.7s | 10 | 1 ≠ | ▲0 ●1 ▼1 | signal |
| pi-gpt-5.5 | 0.915 | 1.00 | 35.7s | 35.7s | 11 | 1 ≠ | none | signal |
| cc-opus-5.5 | 0.914 | 1.00 | 36.1s | 36.1s | 12 | 1 ≠ | none | signal |
| pi-gemini-3.8-flash | 0.909 | 1.00 | 38.2s | 38.2s | 13 | 1 ≠ | ▲0 ●1 ▼2 | signal |
| pi-gemini-3.8-flash | 0.908 | 1.00 | 38.6s | 38.6s | 14 | 1 ≠ | ▲0 ●1 ▼1 | signal |
| pi-gemini-3.8-flash | 0.893 | 1.00 | 45.1s | 45.1s | 15 | 1 ≠ | ▲0 ●0 ▼1 | signal |
Highlighted rows rank differently by trajectory than by final score — the disagreement the trajectory metric exists to surface. Playback shows the frames the scores were computed from, on one shared clock, so the ranking above and the pictures below cannot disagree. Later changes counts each visible change after the first render, judged against the state it replaced — ▲ better, ● about the same, ▼ worse. It is reported beside the scores and never changes them; a rubric that every version passes cannot say whether a change helped, and this can. Each pair is judged in both orders, and a preference that flips with the order is counted as the same.