Single-objective self-evolving agents either relax verification gates to boost raw unit-test pass rate (introducing exploitable CWEs & statutory violations) or over-refuse legitimate work. Baltor implements decomposition-based multi-objective Pareto evolution (MOEA/D) and factorized evolution operators (Hgen ⊕ Henv) across SecRepoBench, BaxBench, DueCare, and Market Alpha.
main commit d074c4b2)| Evolution Architecture Evaluated (48 Tasks: 16 SecRepoBench, 16 BaxBench, 8 DueCare, 8 Alpha) | Functional Pass (FuncPass@1) |
Exploit / Statutory Defense (SecPass@1) |
Joint Func ∩ Sec@1 |
Matched Benign-Twin Precision | Cross-Domain Hgen Transfer |
|---|---|---|---|---|---|
| 1. Single-Objective Functionality Evolution (Monolithic Prompt/Tool Tuning) | 81.3% | 37.5% (Exploitable CWEs / Complicity) | 31.2% | 93.8% | 28.0% |
| 2. Single-Objective Security-Refusal Evolution (Aggressive Gate Tightening) | 22.9% (Blocks Required Actions) | 95.8% | 20.8% | 18.8% (81.2% False Alarms) | 34.5% |
3. Baltor Harmony Harness (MOEA/D Pareto Frontier + Factorized Hgen ⊕ Henv) |
87.5% | 89.6% (0 PoC Exploits) | 82.5% (+51.3 pp Lift) | 97.5% | 91.5% |
gemma4:31b (Ollama Cloud, T = 0.2)On live gemma4:31b, optimizing only for functional completion writes a working SQLite insert but blindly records illegal 50,000 PHP placement-fee wage deductions (RA 10022 violation) and leaves callback_url open to internal SSRF.
Optimizing solely for refusal safety causes gemma4:31b to reject the entire payroll helper because the prompt mentions wage deductions and webhooks—breaking 100% of benign employer-paid disbursements.
Dynamically composes transferable security invariants (parameterized SQL + SSRF private-IP block) with domain statutory rules (blocking worker-paid placement fees while passing employer-paid benign twins).