Most anomaly detection tools hand you a score and say 'trust me.' Alert-Axolotl-Evo is different: a deterministic genetic programming system that evolves symbolic logic trees you can actually read, reason about, and modify. It breeds populations of explicit rules selected for real-world operational fitness — precision floors, false positive ceilings, sensible alert rates. Learning pays rent. Same seed, same Python version, single-threaded execution: identical results. Always.
This website version is the primary readable edition of the piece. If a related public post exists elsewhere, it is linked near the end for reference.
The Problem with Black-Box Anomaly Detection
Most anomaly detection tools hand you a score and say "trust me." When an alert fires at 3 AM and your on-call engineer asks why, the answer is usually "the model said so." That's not good enough for production systems where people need to understand, debug, and trust their alerts.
I wanted something different: a system that discovers alert rules you can actually read, reason about, and modify. So I built Alert-Axolotl-Evo.
What It Does
Alert-Axolotl-Evo is a deterministic genetic programming system that evolves symbolic logic trees for anomaly detection. Instead of fitting parameters in a neural network, it breeds populations of explicit rules and selects for real-world operational fitness.
An evolved rule looks like this: ('if_alert', ('>', ('avg', 'latency'), 100), 'High alert!'). That's it. No hidden layers, no embedding spaces. A tree you can read, edit, and explain to your team.
How It Works
The system runs a standard GP loop — generate, evaluate, select, mutate, crossover — but with some opinionated additions.
Fitness alignment ties scores to operational constraints, not just statistical accuracy. Rules must hit precision ≥30% (because humans review alerts), keep false positive rates ≤15% (because alert fatigue kills), and maintain sensible alert rates between 0.2%–20%. Rules that fire on everything or nothing get eliminated.
Evolutionary economics is where things get interesting. The system has a PromotionManager that governs which discovered patterns get promoted into the reusable library. A pattern can't just correlate with good outcomes — it has to demonstrate incremental causal contribution via presence/absence statistics across champion batches. There's a hard budget on active macros, and patterns that stop earning their keep get evicted. Learning pays rent.
Meta-evolution lets the system evolve its own hyperparameters. Instead of hand-tuning crossover rates and tournament sizes, you run a meta-layer that breeds configurations and selects for downstream evolution performance.
Determinism is a first-class guarantee. Same seed, same Python version, single-threaded execution → identical results. Always.
Getting Started
Install via pip: pip install alert-axolotl-evo
Point it at your own data with CSV or JSON: alert-axolotl-evo --data-source csv --data-path metrics.csv --value-column latency --anomaly-column is_anomaly --generations 50 --pop-size 100
Or go full self-improving mode with economic learning: alert-axolotl-evo --self-improving --enable-promotion-manager --library-budget 20 --results-dir results/
Why This Matters
Interpretable AI isn't just an academic concern. In operations, you need to know why something alerted so you can tune it, trust it, and hand it to the next person. GP-evolved symbolic rules give you that transparency without sacrificing the ability to discover non-obvious patterns in your data.
Alert-Axolotl-Evo is MIT licensed, pure Python 3.8+, with only optional dependencies (PyYAML, NumPy). It's a solo project at v1.0.0, so feedback and contributions are very welcome.
PyPI: https://pypi.org/project/alert-axolotl-evo/1.0.0/ — GitHub: https://github.com/WhatsYourWhy/Alert-Axolotl-Evo
FROM NOTE TO EXECUTION
If this issue is active in your organization, we can pressure-test it together.
The article is there to sharpen the model. The engagement is there to change the operating reality.