Head architecture advances — Qovaryx 3.0

Between the beta1 public build in June and the 3.0 milestone in July, the Qovaryx specialist cluster went through a level of restructuring we hadn't attempted before. The count went up. The authority went down. Both were the point.

The core change: authority tiers, not just heads

In 2.5, every specialist head that scored a signal was, implicitly, allowed to influence an order. That was fine when we had a small handful of heads. It stopped being fine as the cluster grew, because a head with good in-sample metrics but weak out-of-sample transfer could silently distort dispatch.

3.0 introduces an explicit four-tier authority taxonomy. Every head is classified into one of these tiers — and the tier is what determines whether a head can influence a live order, block one, or only be logged for telemetry:

• Scoped execution authority. May affect only a named consumer strategy after out-of-sample testing on real market data. Authority never transfers automatically to a different strategy or order mutation.
• Synthetic shadow. Passed a bounded synthetic-data evaluation but lacks sufficient real forward evidence. It can be logged and compared with the live decision, but cannot mutate an order. Promotion requires a fresh review; observation count alone is not enough.
• Forecast context. Produces a bounded forecast such as volatility or breakout context for an approved consumer. The model itself cannot dispatch an order.
• Offline research. Evaluated outside the order path. It may appear in research telemetry, but has no execution or veto authority.

Nothing overrides an order without a declared head↓consumer↓mutation contract. That contract is versioned in code and enforced by the consumer and shared risk layer.

What changed in the training tempo

During the 3.0 research cycle we ran many head hypotheses through a batch-graded testing pipeline. The common funnel was:

1. CPU screen on a fixed feature snapshot. Cheap screens reject weak ideas before expensive training.
2. Paired-lift test against the current reference for that task. A replacement has to add useful, non-duplicative lift.
3. Consumer-space replay: replay the model through the strategy and order consumer that would actually use it, including applicable costs and abstention behavior.
4. Regime-conditioned stress: does the head hold up in bull, bear, high-vol, and low-vol regimes? A head that's phenomenal in one regime and neutral-to-bad in the others gets scoped to only fire in its regime.

A model's final role depends on the evidence it earned: scoped execution, shadow, context, or offline research. Rejected artifacts are archived off the active model path with their failure reason.

The rehab queue

A subtler advance: we now formally rehab ideas that failed the first time. The method is deliberate: identify the failure mode (feature coverage, regime overfit, label leakage, or consumer mismatch), rebuild only when there is a specific corrective hypothesis, and re-run the full funnel. A replacement inherits none of the prior artifact's authority.

The retirement discipline

3.0 also retired heads. Specifically: heads whose out-of-sample edge collapsed once we finished a proper leakage audit; heads whose training data turned out to be contaminated by split-adjustment glitches; heads whose signal only worked in synthetic labels and evaporated on real-chain replay. Every retirement is dated and reasoned in the internal registry.

The shipping cluster is the survivor set, not a museum of every model that once posted a promising metric. The rejected set matters because it prevents old claims from silently returning to the active path.

The wiring around the heads

3.0 also connected the surrounding safety system: earnings and event gates, options-flow disagreement handling, regime context, feature-coverage checks, broker capability checks, volume/gap prefilters, and symbol exclusions. Packaging now enforces bundled assets and per-user runtime state so development fallbacks cannot impersonate a clean release.

The best head architecture isn't the one with the most heads. It's the one with the most heads that earned their spot and the most retired heads whose failures you can point to.

The next post covers the data moat — the corpora we built to grade all of this on.