[LEVI] Optimize evolutionary optimization architecture

I've been working on making LLM-guided evolutionary optimization (the AlphaEvolve/FunSearch paradigm) cheaper and more accessible. The result is LEVI. The core thesis is simple: most frameworks in this space assume frontier model access and build their search architecture around that. I think this is backwards. If you invest in the harness (better diversity maintenance, smarter model allocation) you can get the same or better results with a 30B model doing 90%+ of the work. Two ideas make this work: **Stratified model allocation.** Cheap models (Qwen 30B) handle most mutations. Expensive models only get called for rare paradigm shifts where you actually need creativity. The evolutionary process is blind anyway. FunSearch reached their capset result with a \~30B model over a million mutations. Raw model intelligence isn't what drives the breakthroughs, compounding blind search is. **Fingerprint-based CVT-MAP-Elites.** Instead of choosing between structural diversity (OpenEvolve) or performance-based diversity (GEPA's Pareto fronts), we use both as dimensions of a single behavioral fingerprint. Centroids are initialized from structurally diverse seeds with noise perturbation, so the archive doesn't overfit to early strategies or waste space on regions no program will ever visit. **Results:** On the UC Berkeley ADRS benchmark (7 real-world systems problems: cloud scheduling, load balancing, SQL optimization, etc.): |Problem|LEVI|Best Competitor|Cost Savings| |:-|:-|:-|:-| |Spot Single-Reg|51.7|GEPA 51.4|6.7x cheaper| |Spot Multi-Reg|72.4|OpenEvolve 66.7|5.6x cheaper| |LLM-SQL|78.3|OpenEvolve 72.5|4.4x cheaper| |Cloudcast|100.0|GEPA 96.6|3.3x cheaper| |Prism|87.4|Tied|3.3x cheaper| |EPLB|74.6|GEPA 70.2|3.3x cheaper| |Txn Scheduling|71.1|OpenEvolve 70.0|1.5x cheaper| LEVI also beats AlphaEvolve's circle packing score while mostly using Qwen 30B. The part I think is most interesting is the controlled comparison: same model (Qwen3-30B-A3B), same budget (750 evals), three seeds. LEVI reaches scores within 100 evaluations that neither OpenEvolve nor GEPA hit at any point. So the gains come from the search architecture, not just throwing a bigger model at it. Blog: [ttanv.github.io/levi](https://ttanv.github.io/levi) Code: [github.com/ttanv/levi](https://github.com/ttanv/levi) Happy to discuss the architecture, diversity mechanism, or cost breakdown. Sorry for the repost, used the wrong flair last time.