The Specialist Committee: Why We Split AI Into Six Models Instead of Building One

The intuitive approach to a prediction problem is to build one model, give it all the features, and let it learn.

This approach fails in a specific and predictable way: it learns something about everything and a lot about nothing. The gradient descent process allocates capacity across the entire feature space simultaneously. Signals from different domains — some of which are relevant on short timescales, some on long timescales, some conditionally, some continuously — compete for the same representational capacity. The model averages across these timescales rather than specialising in any of them.

The alternative is a committee of specialists.

The Architecture

The specialist committee approach assigns each model a single domain. It is trained exclusively on features from that domain. Its objective is to be the most accurate possible predictor within that domain, not across all domains simultaneously.

In the market microstructure research programme that informs the TVS analytical methodology, the committee comprises six specialists:

Wall structure — the behaviour of large resting orders in the order book: their appearance, decay, reinforcement, and relationship to subsequent price movement. The wall specialist learns the signature of genuine institutional support versus spoofed liquidity designed to mislead.

Aggression — taker-side behaviour: the rate, size, and clustering of market orders. Sustained aggressive buying has a different signature to a single large sweep. The aggression specialist learns to distinguish them.

Order flow — the net directional pressure of executed trades over multiple time windows. Cumulative delta, flow persistence, flow reversal. The flow specialist learns which flow configurations precede continuation versus exhaustion.

Smart money — large-participant activity: block trades, unusual concentration of size, the timing of institutional entries relative to the order book structure they are trading into. The smart money specialist learns the fingerprint of informed participation.

Momentum — price-based and volume-weighted trend signals across multiple horizons. Persistence, acceleration, the rate at which momentum decays. The momentum specialist learns which momentum configurations are durable versus fragile.

Context — the broader market environment: session timing, volatility regime, cross-market signals, macro context. The context specialist provides the conditional frame within which the other specialists’ outputs should be interpreted.

What the Validation Revealed

Training the specialists independently and then evaluating their cross-horizon information coefficients produced a finding that would not have been visible in a monolithic model: the specialists generalise very differently across timescales.

The wall structure specialist, trained to predict outcomes at one timescale, remains predictive at timescales two to three times longer. The signal it captures — the behaviour of large resting institutional orders — carries information that persists across timescales because institutional programmes run on longer cycles than individual trades.

The flow and aggression specialists show the opposite pattern: their predictive signal decays rapidly as the evaluation horizon extends beyond their training horizon. Taker aggression is a short-window phenomenon. What the market is doing in the next 30 seconds is strongly reflected in current taker behaviour; what it will do in the next three minutes is much less so.

This distinction — which specialists generalise and which are domain-constrained — is not available in a monolithic model. The model learns an average. The specialist committee makes the heterogeneity visible, which means it can be engineered for.

A downstream routing architecture can therefore use the wall specialist’s signal at long horizons and the aggression specialist’s signal only at short horizons — with confidence that this is not arbitrary weighting but an empirically supported separation.

The Meta-Layer

Above the six specialists sits a meta-routing layer.

The meta-router does not simply average the specialists’ outputs. It learns which specialists are reliable under which conditions, and routes confidence accordingly. In periods where the wall structure is ambiguous — where genuine and spoofed liquidity are difficult to separate — the wall specialist’s contribution is downweighted. In clean market conditions where the wall structure is clear, it is upweighted.

The ensemble of specialists with a meta-router is a materially different architecture from either a single monolithic model or a naïve average of individual models. The monolithic model cannot introspect on its own domain reliability. The naïve average treats all specialists as equally valid at all times. The meta-routing layer learns the conditional reliability landscape.

The result — approximately forty trained model checkpoints, with the meta-router sitting above them — is what we call the God Committee in the research documentation: not because it is infallible, but because it is the most complete attempt to fuse specialist knowledge that the research programme has produced.

The Two-Engine Production Architecture

The specialist committee produces scores. The production system that operates those scores runs on a two-engine architecture.

The first engine is a structural detector. It is a shape-matcher: it identifies when the current market state matches a known candidate pattern — the multi-stage temporal signatures discovered in the research phase. A background condition that has been building over tens of minutes. A formation condition in the intermediate window. A trigger event in the final seconds. The structural engine detects when all three are present simultaneously.

The second engine is a context evaluator. When the structural engine raises a candidate, the context engine does not immediately act on it. It asks: are the conditions under which this pattern has historically been reliable currently present? If the context engine finds that the broader market environment is misaligned with the candidate pattern’s historical reliability conditions, it suppresses the candidate.

Only signals that pass both engines propagate.

The two-engine design exists because structural pattern recognition and contextual reliability are fundamentally different problems. A structural detector that does not include a context check will act on every valid shape, including the shapes that historically have not held up when broader conditions are adverse. A context evaluator without a structural detector has nothing concrete to evaluate.

Separating them and running them in sequence produces a system that is selective rather than reactive — one that declines to act when the evidence is structurally valid but contextually unsupported.

The Warm-Start Governance Principle

There is an architectural discipline in the production system that deserves specific attention.

The context evaluation engine does not begin processing candidates until the quality of its input data is verified. If a supplementary data source is stale or unavailable, the context engine waits. It does not operate on degraded inputs.

This is easy to state and difficult to maintain under production pressure. The bias in any live system is always toward availability — toward returning a result, any result, rather than declining to operate when inputs are inadequate. A context engine that produces output on degraded data looks functional. It is not. It is generating plausible-looking conclusions from insufficient evidence.

The warm-start governance principle — the structural refusal to operate until inputs are verified as clean — is the difference between a system that is reliably accurate when conditions are good and a system that generates output of unknown quality under all conditions.

The Connection to TKAD

The specialist committee and the warm-start governance principle are not limited to financial market microstructure. They are architectural choices about how to build AI systems that know what they don’t know.

TKAD — the TVS methodology for knowledge absorption and distribution — applies both principles at the pipeline level.

The absorption pipeline is not a single model processing everything. It is a sequence of specialised functions: ingestion by source type, entity resolution, epistemic classification, provenance tracking, privacy transformation, semantic chunking, retrieval. Each function is a specialist. Each has a defined input, a defined output, and a defined domain.

The human-in-the-loop verification layer is the warm-start governance principle applied to knowledge publication. The distribution engine does not operate on unverified knowledge. It waits. It does not publish because it can; it publishes only after the human authority over the knowledge has confirmed that what the system has extracted is accurate, current, and appropriate for distribution.

The same discipline that killed Signal 1 in the market research — building infrastructure that refuses to propagate what it cannot verify — is the discipline that governs what gets published under TKAD.

The system that finds and kills its own false positives is the same kind of system as the one that refuses to distribute unverified knowledge. The architectural principle is identical. The domain is different.

---

The Signal 1 retraction is at The Most Valuable Thing We Built Wasn’t a Signal. The staged discovery methodology is at The Correlation Engine. The evolutionary signal search that found the surviving candidates is at How We Search for Signals We Haven’t Imagined Yet. The market physics reconstruction that produced the episode taxonomy is at Market Physics: Why We Model Behaviour Before We Model Price. The production crash history that proves eighteen months of live operation is at The Crash Diary. The TKAD pipeline that applies these architectural principles to knowledge distribution is at TKAD: Engineering a System for the Knowledge AI Cannot Invent.