# E72 — Limitations and non-claims

**Measured, not assumed**
- No remote model provider was executed. 0 remote calls; three local models only (135M/362M/494M, CPU).
- Cross-host execution is real (two VMs, two mechanisms: SSH/SQLite and PostgreSQL). Coordinator failover is
  demonstrated on a repeated run (host B took leadership, fencing 2->3, M2 COMPLETED, 0 duplicates); the first attempt
  failed and is recorded. Network partition was **not** induced. PostgreSQL is a single primary, not Byzantine fault
  tolerant.
- A full host reboot **was demonstrated once** on development replica B (boot id changed, uptime reset, mission M4
  completed 100/100, expired grants refused). This is not a repeated soak or a production DR proof. The separate
  SSH/SQLite demo did not attempt a reboot.
- The long-horizon campaign is 1 hour, duty-cycled. 6/24/72/168-hour missions and 500/1,000/5,000/10,000 verified
  productive-hour targets are **not achieved**.
- Benchmark samples are small; the local model is weak; deterministic families are not comparable to model families.
- The repository mission's pre-fix variant reconstructs an observed E72 defect; it is not a pre-existing HEAD bug.
- The security target was written for the exercise; the lab is same-project.
- ClawX is local and not deployed; the evidence key is self-generated (internal consistency, not third-party identity).
- No third-party audit, no SOC 2, no certification. The team remains Vincent Noriega (founder), Lena Noriega and
  Monica Diaz.

**Never claim from this work**
"production-ready", "independently audited", "hardware-attested", "10,000 productive hours", "24/72/168-hour
mission", "network partition tolerance", "repeated reboot recovery", "remote model capability", "general intelligence",
"notarized/court-admissible evidence".