Evolution 72 — autonomous intelligence, distributed execution, real-world endurance
Everything below is measured from raw records in the signed bundle. No third-party audit, no SOC 2, no machine-reboot recovery, no cross-host failover, no remote model call. Those are marked BLOCKED or PARTIAL, not claimed.
Status: 21 PASS · 4 PARTIAL · 3 BLOCKED · 0 FAIL. Final HEAD 3fa6c6fb826e6a51fa74cb7af51caaaaefb6c7d3 on evolution-72.
Measured facts
| Inherited baseline | E68 1018.29 worker-h; E70 163.4645 h; E71 6.65 h. All inherited suites pass. |
|---|---|
| E70 discrepancy | Preserved both: 163.4645 h (alive at measurement) vs 163.894 h (all rows). Delta explained by 3 replaced workers. |
| Scheduling | 6 concurrency levels, controller lease expiries 0; naive 1 s lease failed with 36 expiries. |
| Models | 56 real local calls; 0 remote calls. Accuracy: qwen2.5-0.5b 4/8, smollm2-135m 0/8, smollm2-360m 1/8. |
| Missions | 5/5 accepted: repository patch + independent failing/passing test, security exploit/fix, arXiv research, Iris data, host reliability. |
| Cross-host | Two VMs. SSH/SQLite: {"host-a": 18, "host-b": 22} verified tasks, 0 duplicates. PostgreSQL: M1 COMPLETED, 8 cross-host verified, 0 duplicates, 155-event chain ok=True. Failover: first attempt failed; repeated run took over (b_took_over=True, fencing 2->3, M2 COMPLETED). Network partition not induced. |
| Recovery | Worker kill: COMPLETED; DB lock fails closed: True; cold restart: COMPLETED. Real host reboot demonstrated on host B: boot_id changed=True, uptime reset=True, M4 16/16 verified, 6 expired grants refused. |
| Long horizon | 172 cycles, 688 verified EXECUTE tasks, 3618.66 s elapsed, 0.0174 productive h, mode DUTY_CYCLED. 6/24/72/168 h are BLOCKED. |
| Memory | Verified retrieval 1.0 vs no memory 0.0; injection quarantined; authority violations 0. |
| Self-improvement | model_routing +0.25; scheduling_admission +0.6666666666666667; promotion AWAITING_OWNER_APPROVAL. |
| Benchmark | Local model 0.44 exact on 25 calls; deterministic families 5/5, 5/5, 5/5, 20/20; GSM8K 1/10. |
| Economics | 4 workers util 0.8132 / peak PSS 25.6 MB; 8 workers util 0.4614 / 46.6 MB. |
| Security lab | 103/103 scenarios blocked, 0 bypassed. Same-project lab, not an audit. |
| Gates | 21 PASS / 4 PARTIAL / 3 BLOCKED / 0 FAIL. |
Download and verify
- Signed E72 bundle manifest
- Pinned evidence key
- Standalone verifier (no CAIN imports)
- Final audit (Markdown)
- Final audit (JSON)
- Gate registry
- Full archive (.tar.gz)
- SHA-256 sums
- History: prior E72 audit (1c928a4ca)
- Mission evidence
- Recovery evidence
- Long-horizon evidence
- Documentation
- Tests
python3 verify_e72_offline.py bundle --jsonThe verifier checks the Ed25519 manifest signature against the pinned key, recomputes every file digest, checks gate evidence presence, re-derives the arithmetic claims, and verifies the embedded v5 mission bundles. The key is self-generated: it proves internal consistency and tamper detection, not a third-party identity. Pin the fingerprint out of band if you need that.
Not claimed
- No production readiness, third-party audit, SOC 2 or certification.
- Host reboot was demonstrated once on host B (boot_id changed, M4 100/100); repeated reboot/soak recovery is not claimed.
- No network partition was induced. Cross-host failover succeeded on a repeated run; the first attempt failed and is recorded in the docs.
- No remote model provider: three local models only (135M/362M/494M parameters, CPU).
- No 6/24/72/168-hour mission and no 10,000 productive-hour target.
- No general intelligence; benchmark samples are small and carry Wilson intervals.