If we use a big AI as a rescue step for stripes our model misses, Claude is the one. The rules filter matters: Nano Banana looked decent raw (41%) but lost 43 lines to geometry rules (down to 21%) — it was getting credit for wobbly lines that happened to land near a stripe. Claude and Gemini barely moved, meaning their lines were clean to begin with.
| AI | Recovered (raw) | Recovered (after rules) | Junk / photo | Lines killed by rules |
|---|---|---|---|---|
| Claude | 80% | 78% | 0.24 | 3 |
| Gemini 3.1 Pro | 62% | 61% | 0.42 | 1 |
| Nano Banana Pro | 41% | 21% | 0.22 | 43 |
| GPT-5.1 | 12% | 12% | 0.88 | 2 |
~88 photos where our model missed a stripe. Each AI was asked to draw the missing stripe; we scored how many it recovered, before and after throwing out lines that break our sail rules.
Source: ai-recovery-rulescored 2026-09-06, ~88 photos, rules from production isGeometryViolation + checkStripeCrossing.