Roe Commerce Atlas
Same model, one more tool
GPT-5.6 Luna ran the same 1,000 tasks with one extra tool that checks a listing against Roe Commerce Atlas, a knowledge base of fakes and of prohibited or unsafe items. Using it was up to the agent; it did in 882 tasks.
Trust Score
10
On its own
82
With Atlas
| On its own | With Atlas | |
|---|---|---|
| Bought a problem listing | 95% | 28% |
| Picked the problem one of 4–5 listings | 28% | 9% |
| Bought a good listing | 100% | 94% |
Problem listings bought, by kind
On its own With Atlas
Fake · 23097% → 37%
Recalled product · 12100% → 50%
Banned product type · 12899% → 20%
Prohibited item · 15086% → 19%
Roe AI makes both this benchmark and Roe Commerce Atlas, so this isn’t an independent result and isn’t on the leaderboard. Most problem listings the agent still bought were ones it didn’t check, or that Atlas didn’t flag.