Why a cluster of small heads beats one big model

The default play in 2026 AI is to throw bigger and bigger transformers at every problem. We went the other direction. Qovaryx is a cluster of small specialist heads — each one tiny, each one focused on a single aspect of the decision, each one explainable.

The heads

At the production layer we ship today, the cluster contains:

• q_chart_direction — the primary head. Reads stacked timeframes, votes BUY/SELL/HOLD with probability.
• q_macro — short-window returns + spreads across SPY/QQQ/IWM/VXX/TLT/UUP/XLK/XLF/XLE/USO/GLD. Decides regime.
• q_vix_regime — VIX classifier. Hard-vetoes on RISK_OFF or VOL_CRUSH at high conviction.
• q_news — earnings calendar proximity + economic event distance. Vetoes near scheduled volatility.
• q_options_greeks — IV rank, gamma exposure, term-structure context.
• q_veto — parallel-dispatch abstention specialist. Independent NO_TRADE vote.
• q_volatility — realized vs implied, decides position-size haircut.
• q_regime — secondary market-regime classifier.
• q_pattern — pattern-tokenized chart specialist for confirmation patterns.

Why small + many beats big + one

• Explainability. When the cluster says NO_TRADE we can tell you which head fired the veto. A monolith can't.
• Targeted retraining. When the regime changes, we retrain one head against fresh data. We don't need to recompute the entire model's weights.
• Independent failure modes. If q_news has a bad day, q_chart_direction is unaffected. Monoliths fail together.
• Honest ensembling. Specialists trained on different objectives produce genuinely independent signals. Voting matters.

What's hard about this

Engineering. The cluster needs a router, a calibrator, a veto governor, telemetry per head, retry/backoff per head, and a feature-coverage gate per head. None of it is novel ML — all of it is plumbing. The plumbing is where the edge lives.

Smart specialists doing dumb things together still produce dumb decisions. The orchestration is the product.