Benchmarks & SLOs
Exact where code is exact. A range where the world varies.
Deterministic guarantees and provider-dependent measurements are published separately. Combining them would overstate what the data can prove.
Reproducible-exact
- Every delivered artifact carries its Tribunal report: true.
- No hard check can fail inside a passing pack: true.
- Undelivered work is declared, never silently dropped: true.
Measured-with-variance · 14 real runs
| tool | n | median | range | pass median |
|---|---|---|---|---|
| oce_write_toast | 3 | 12.1s | 9.4–13.7s | 100% |
| oce_plan_occasion | 3 | 105s | 96.9–107.9s | 100% |
| oce_moodboard | 2 | 56.5s | 36.8–76.1s | 100% |
| oce_make_keepsake | 2 | 100.5s | 78.5–122.5s | 100% |
| oce_design_invite | 2 | 70.4s | 68.6–72.2s | 100% |
| oce_launch_kit | 2 | 337.8s | 194.3–481.3s | 75% |
Measured 2026-07-14 with $2.61 in real provider spend. Reproduce with node scripts/slo.mjs.
Small samples remain labelled small. The fuller explanation, including critic variance, is on the public evaluation page.