# E111 Limitations

1. **Wall clock.** The six-hour gate depends on the actual run. At bundle build the state is
   `DONE` and the gate is `PARTIAL`.
2. **Same project.** All experiments were designed, run and verified inside this project. No third party has
   reproduced them.
3. **Constructed intelligence tasks.** Skills, transfer and collective tasks are generated with deterministic
   oracles. They measure procedures, not broad competence.
4. **Deterministic specialists.** The collective is code, not independent model agents.
5. **No neural learning.** E111 changes procedures and memory, not weights.
6. **One administrative domain.** The remote host is a separate machine but not a separate administration.
7. **Local sources.** Real-world evaluation uses public datasets and the project's own source/operations data;
   there is no field deployment or customer system.
8. **Small samples.** Skill sets have 26-43 held-out cases; exact tests have low power.
9. **No production MCPGate.** Enforcement evidence is loopback plus one remote one-shot resource.
10. **Promotion withheld.** The self-improvement candidate is not enabled.
