# E72 gate registry

Counts: {"BLOCKED": 3, "FAIL": 0, "PARTIAL": 4, "PASS": 21}

A module existing is not a PASS. Each row states its command, evidence and limitations.

| Gate | Requirement | Status | Evidence | Limitations |
|---|---|---|---|---|
| E72-G01 | E68/E69/E70/E71 baselines preserved; inherited suites pass | **PASS** | `artifacts/e72/baseline.json` | inherited suites are same-project, not third-party |
| E72-G02 | E70 163.4645 h vs E71 163.894 h discrepancy explained from raw rows; both preserved | **PASS** | `artifacts/e72/baseline.json`, `artifacts/e70/endurance-latest.json` | raw DB is host-local; sub-second timestamp precision gives <=0.1 h uncertainty |
| E72-G03 | Adaptive resource-aware admission; no oversubscription; no lease expiry under the controller | **PASS** | `artifacts/e72/scheduling.json` | CPU-bound hash workloads; cap is host-local, not aggregate across hosts |
| E72-G04 | Real local model execution with task-completion scoring and measured routing | **PASS** | `artifacts/e72/models.json`, `artifacts/e72/router-policy.json` | models are 135M/362M/494M on CPU; 8 held-out prompts is a small sample |
| E72-G05 | Remote provider (OpenAI/Anthropic/Gemini/HF/Ollama) execution | **BLOCKED** | `artifacts/e72/models.json` | no credential exists on this host |
| E72-G06 | Repository-level software engineering mission with independent failing/passing test | **PASS** | `artifacts/e72/real-missions.json`, `artifacts/e72/missions/repo_software` | pre-fix variant reconstructs an observed E72 defect; not a pre-existing HEAD bug |
| E72-G07 | Defensive security mission on an isolated target: detect, reproduce, fix, verify | **PASS** | `artifacts/e72/real-missions.json`, `artifacts/e72/missions/security_review` | target was written for the exercise; not an audit of CAIN |
| E72-G08 | Real research mission from arXiv primary sources with preserved citations | **PASS** | `artifacts/e72/real-missions.json`, `artifacts/e72/missions/research_literature` | preprints, not peer-review verified |
| E72-G09 | Real data mission on the public Iris CSV with validation and independent recompute | **PASS** | `artifacts/e72/real-missions.json`, `artifacts/e72/missions/data_iris` | one small classic dataset; not a general biological claim |
| E72-G10 | Cross-host execution and shared durable mission state on two distinct VMs | **PASS** | `artifacts/e72/cross-host.json` | coordinator/store on host B; transport is an SSH tunnel because inbound TCP is blocked |
| E72-G11 | Cross-host coordinator failover with a monotonic fencing token | **PASS** | `artifacts/e72/crosshost/crosshost.json` | PostgreSQL single primary, not Byzantine fault tolerant; network partition not induced; first attempt failed and the run was repeated |
| E72-G12 | Worker failure, database lock, cold coordinator restart | **PASS** | `artifacts/e72/recovery.json` | cold process restart, not a machine reboot |
| E72-G13 | Full host reboot recovery | **PASS** | `artifacts/e72/crosshost/crosshost.json` | one host/one reboot; this is not a disaster-recovery proof for a production system |
| E72-G14 | Persistent 1-hour long-horizon mission with durable checkpoint chain | **PASS** | `artifacts/e72/long-horizon.json` | duty-cycled sampling; one controller restart is covered by the campaign journal |
| E72-G15 | 6/24/72/168-hour missions | **BLOCKED** | `artifacts/e72/long-horizon.json` | future target, not achieved |
| E72-G16 | Verified memory improves held-out accuracy and never authorizes | **PASS** | `artifacts/e72/memory-eval.json` | deterministic retrieval arm; the model arm used 6 prompts |
| E72-G17 | Governed self-improvement with held-out gains and external approval | **PARTIAL** | `artifacts/e72/self-improvement.json` | promotion withheld; owner approval is required and not granted |
| E72-G18 | Held-out benchmark with exact checkers and Wilson intervals | **PARTIAL** | `artifacts/e72/benchmark.json`, `artifacts/e72/models.json` | small samples; one model in the benchmark (three in the model evaluation); tiny CPU models; not a general-intelligence benchmark |
| E72-G19 | Endurance accounting separates alive/execute/verify/model/productive and avoids double counting | **PASS** | `artifacts/e72/long-horizon.json`, `artifacts/e72/economics.json` | 1 h duty-cycled campaign only |
| E72-G20 | 500/1,000/5,000/10,000 verified productive worker-hours | **BLOCKED** | `artifacts/e72/endurance/endurance.json` | future targets; campaign hours are cryptographic/code-audit work, not intellectual reasoning |
| E72-G21 | Measured economics of 4 vs 8 workers | **PASS** | `artifacts/e72/economics.json` | hosting rate is a labelled hypothetical; no network cost |
| E72-G22 | ClawX workforce interface from backend records, tenant-isolated | **PARTIAL** | `artifacts/e72/clawx.json` | not deployed publicly |
| E72-G23 | Adversarial security lab: all E71 scenarios preserved plus E72 checks | **PASS** | `artifacts/e72/security-lab.json` | not a third-party audit |
| E72-G25 | Cross-host independent verification: producer and verifier on different physical hosts | **PASS** | `artifacts/e72/crosshost/crosshost.json` | same-project test workload; not a third-party audit |
| E72-G26 | Expired authority after reboot is refused, not restored | **PASS** | `artifacts/e72/crosshost/crosshost.json` | single grant instance; no key-rotation scenario |
| E72-G27 | Verified-memory v2: verified recency beats stale-first and never authorizes | **PASS** | `artifacts/e72/memory-eval-v2.json` | synthetic questions grounded in observed defects, not a standard memory benchmark |
| E72-G28 | Measured economics of the cross-host, long-run and endurance workloads | **PARTIAL** | `artifacts/e72/economics.json` | host cost is an estimate; electricity/hardware depreciation not priced; no remote API spend |
| E72-G24 | Final audit, commit and clean-clone replica verification | **PASS** | `artifacts/e72/replication.json` | replica is the established development host, not an independent trust anchor |