Abstract
We hypothesize that some quantitative signals fail not only because they were fitted poorly, but because widely learnable patterns can become crowded or easier to imitate as capital responds. We refer to behavior consistent with that hypothesis as adversarial capture.
This paper outlines ChartChest's architectural response: a self-improving intelligence flywheel that combines concept-first reasoning, orthogonal evidence channels, and a human + AI out-of-sample learning loop. The central claim is architectural, not statistical: durable edge comes from a system that keeps asking better questions, not from a model that answers one question well.
1. The Problem
The default market-ML loop is fit, deploy, and repeat. In adaptive markets this is fragile, because profitable templates are eventually exploited. We observe a recurring pattern: the setups that look most textbook-correct are often the easiest to stage and fade.
| Signal family | Trend across regimes | Interpretation |
|---|---|---|
| Textbook-correct composites | Fades toward zero | Decays as the pattern becomes known |
| High-conviction everything-agrees signals | Fades, then inverts | Most dangerous when most confident |
| Signals expensive to counterfeit | Stable to strengthening | More durable under adaptation |
2. Discovery One: Artificial Wisdom
Rather than optimizing for pattern recall, we anchor on questions bounded by real-world cost and constraints. In plain terms: what is happening, when, who is doing it, where evidence clusters, and why now.
The five words are not the edge. The edge is how each is mapped to observables, how confidence is composed, and how those questions evolve as the adversary adapts. That operational mapping is intentionally withheld.
3. Discovery Two: Concepts over Fragments
In the experiments reported here, shallow split models performed better when they reasoned over curated concepts instead of raw detector fragments. We therefore tested named market theses before asking the decision layer to reason over agreement, contradiction, and uncertainty.
The historical out-of-sample result improved on the tested feature-level baselines. The narrower lesson was that concept architectures depended on calibration at each concept's natural input scale; otherwise concepts silently failed to activate.
4. Discovery Three: Orthogonal Evidence, Cascaded
Multiple specialized reasoners are composed as a cascade rather than a simple average. A new model helps only when it brings a genuinely new channel of evidence. Model count is not the lever; evidence orthogonality is.
We also found that training objective alignment dominates architecture tweaks. A model is only as good as the question its label encodes.
5. Discovery Four: Human + AI OOS Flywheel
The system is built around commitments made before outcomes are known. Later, those commitments are scored by reality and folded back into retraining under a champion/challenger promotion gate.
This loop is designed to build a prospective record over time. Only observations registered before outcomes are known can contribute fresh evidence; elapsed operation does not by itself establish an edge.
6. Methodology
- Time-ordered walk-forward validation with leakage controls.
- Out-of-fold evaluation discipline for all headline claims.
- Promotion gating so only verified improvements ship.
- Direction-of-result reporting over metric theater.
7. What We Learned
- The reported tests found behavior consistent with recurring adversarial capture.
- Constraint-grounded reasoning was more durable than pattern memorization under the tested design.
- Concept-first representation outperformed the declared raw-feature baseline in those experiments.
- Orthogonal evidence and objective alignment were the primary measured levers.
- The out-of-sample loop is an evidence-building discipline, not proof of live profitability.
8. What Is Deliberately Withheld
This paper intentionally omits all reconstruction-critical details: mappings from guiding questions to observables, calibration scales and thresholds, objective decomposition and cascade order, and loop-hardening mechanics. We also withhold exact coefficients and per-experiment internals.
9. Conclusion
Markets can adapt, and predictive relationships can decay. The research program therefore favors better questions grounded in constraints and requires repeated out-of-sample testing rather than assuming a static predictor will remain useful.
Continue reading: Part II — Trying to Break Our Own Edge, a pre-specified falsification program; Part III — Wisdom, Judged, the architecture of the judgment layer; Part IV — Reading the Opponent, the engines that read the tape as authored behavior; Part V — Reading the Move, where direction and execution authority come apart; Part VI — The Context We Preserved, where strategic meaning is consolidated and preserved across the system; and Part VII — The Weave That Holds, the capstone on selective context weaving and a final model that can keep learning without losing its meaning.