[ryansolid]

Adds the write-up and grades for three Opus apps built in Exact (the contract arm from #59), run the way we ran the .rx probe. See eval/results/desk13-exact-2026-11-13/README.md.

• Defects: 1, 2 and 0, a mean of 1.0. The last three .rx apps scored 2, 0 and 0 (mean 0.67). With three apps per arm, neither difference is meaningful.
• Isolation: agents worked from a copy of Exact with bench/ removed, so they could not read Exact's desk9 reference.
• Shared defect: two of the three apps left the settings save on Exact's default mutation mode instead of parallel. A later save then replaces the earlier one, so the earlier result never shows.
• Cost: Exact apps used 145–168k tokens and 10–12 minutes each, against 99–107k tokens and about 5 minutes for .rx.
• Size: 568–605 lines per Exact app, against 411–452 for .rx.

---

Generated by Claude Code