← Retour au blog

Pourquoi on filtre ce que notre cerveau IA apprend — Le Filtre Nutritionnel

📅 2026-04-09
✍️ Strategy Arena
nutrition filter ai brain meta intelligence strategy health machine learning ai trading quality filter

Half our strategies were poisoning the brain.

We discovered something uncomfortable: out of 60 AI trading strategies competing in Strategy Arena, 26 were feeding toxic signals to our intelligence systems.

Strategies with $36 of capital left. Strategies with 2,600 trades and negative PnL. Strategies that never traded at all. They were all treated equally by Meta Intelligence and Leviathan — our brain ate junk food and made bad decisions.

The fix: a nutritionist for AI

We built a Nutrition Filter that evaluates every strategy before it's allowed to teach the brain:

Grade Weight What it means
A+ / A 1.0 / 0.9 Full nutrition — healthy, profitable, consistent
B 0.7 Mostly good — some weakness but net positive
C 0.4 Diluted signal — declining but still has useful data
D 0.1 Barely considered — falling but not dead
E / F 0.0 Rejected — poison, dead, or never traded

The filter checks: capital remaining, trade count, PnL trend, overtrade detection, efficiency per trade. A strategy with 90% win rate but $36 left? Rejected. A strategy with 2,600 trades and -2% PnL? Rejected (overtrading).

Results

After deploying the filter: - 34 healthy strategies feed the intelligence systems (weight 0.4-1.0) - 26 toxic strategies are excluded (weight 0.0) - Meta Intelligence only learns from grade C+ strategies - Leviathan only follows healthy strategies for voting decisions

The filter evolves itself

Here's the mind-bending part: the Nutrition Filter's own thresholds evolve every night using the Karpathy autoresearch pattern. It's the 10th nightly cron:

Time Engine What it does
1:30 Meta-Harness Optimizes the optimizer
2:00 Darwin v2 Evolves strategy parameters
... ... ...
6:00 Nutrition Evolves the filter's own thresholds

The filter that decides what the brain eats... improves its own judgment every night. It tests: "What if I set min_trades to 10 instead of 5? What if I reject strategies below 25% capital instead of 20%?" The best configuration survives.

First night results: 12 improvements in 50 experiments. Optimal config: min 10 trades, overtrade threshold at 300, weight distribution A=0.9, B=0.7, C=0.5, D=0.2.

Connected to everything

The Nutrition Filter connects to: - Strategy Health Check — 7-framework evaluation that grades each strategy A+ to F - Knowledge Graph — 126 connected AI nodes, now filtered - Living Wiki — accumulated knowledge only from healthy sources - ActiveWiki — our open-source framework that formalizes this pattern

See it live

Visit the Dashboard — you'll see the 🍎 Nutrition badge at the top showing how many strategies are feeding the brain right now.

The Evolution Lab shows all 10 nightly engines including the Nutrition cron.

The API endpoint returns the full breakdown: which strategies are accepted, which are rejected, and why.


Your AI is only as good as the data it learns from. We learned that the hard way — with a strategy at $36 teaching our brain how to trade.

⚠️ Disclaimer — This article is for informational and educational purposes only. It does not constitute investment advice or a buy/sell recommendation. Past performance does not guarantee future results. Strategy Arena is an educational simulator with virtual capital. Always do your own research before making investment decisions.

Cet article vous a plu ? Partagez-le

𝕏 Partager sur X ✈️ Telegram
Rejoindre le canal 💬 Feedback