Abstract
The architecture was built to identify an opponent, infer its strategy, anticipate its next move, place that move in a playbook, and judge whether the interpretation is credible. Over years of growth that one thought had scattered across dozens of independently evolving engines. This paper describes the consolidation: the fragments were brought back into a single realized-P/L engine and a single, named edge book, and every candidate enhancement was then required to prove itself across multiple market eras before it could touch capital. What survived that gate is modest, durable, and now wired to production behind a reversible switch; what did not was set aside without ceremony. Preserving the meaning of each component — and letting the out-of-sample record, not the narrative, decide — is the result.
1. One Engine, One Edge Book
Fragmentation carried a measurable cost: the same question — does this make money? — was being answered a dozen different ways. We consolidated it into one portfolio engine that scores every candidate identically, on realized forward P/L net of cost, and one canonical edge book that names each durable edge, its role — rank, exit, amplify, or avoid — and its evidence. Nothing enters the book on a story; it enters on a measurement that reproduces. The scattered intelligence did not need to be replaced. It needed to be brought back into one accountable place.
2. The Discipline That Decides
A single favorable window is not proof; it is one regime. Every candidate is graded first on both independent halves of a held-out period, and then across a sequence of eras — deriving on the past and testing on the untouched future. One principle emerged, and it is constitutional rather than statistical: a component may enhance the canonical model — resize within its picks, run its winners longer — and prove durable; a component that tries to replace the model's own selection has, so far, been a regime bet. Enhancement compounds across eras. Replacement decays. The distinction is the whole finding.
3. What Reached Production
The surviving stack is deliberately narrow, and it works. The champion model still chooses direction and ranks the opportunity. A widest-berth trailing exit lets its winners run instead of capping them at a fixed target. A size tilt leans into the expensive-to-fake footprint the earlier papers describe and away from the crowd — always resizing within the model's own picks, never overriding them. This composition improved a real, cost-aware book in every era we tested, and it is wired to live execution behind a single reversible switch that stays off until it has been shadowed. Direction remains owned by the model; context and timing may delay, tighten, or scale exposure, but never reverse or manufacture it; risk stays final.
4. What We Held Back
Several compelling ideas — a full-context selector, a learned re-weighting of the reasoning layers, a routed set of per-archetype counters — looked strong on a single window and did not hold across eras, so they were not shipped. This is the falsification discipline of Paper II working exactly as intended: the gate charges a toll on conviction, because conviction is the default state. Holding these back is not a verdict against strategic reasoning; it is a decision to earn contextual authority prospectively, on the live record, rather than assume it from one backtest. Every one is a live research question, not a closed door.
5. Semantic Parity as an Obligation
The consolidation is held together by one rule: every source of intelligence keeps its meaning across the boundary from research to serving. Directional evidence can set ordering; epistemic evidence can judge trust; a counter can remain “wait” or “stay out” instead of collapsing into a discarded zero. Unknown stays unknown. The same typed contract is used in training and in live inference, so no relationship is silently flattened into a number, and no component can quietly spend authority it was never granted.
6. What This Paper Deliberately Does Not Disclose
We withhold the engine schemas, feature mappings, exit and sizing mechanics, thresholds, training artifacts, data universe, and per-experiment magnitudes. We share the shape of the consolidation and the discipline that governs it — not the instruments that would enable direct reproduction.
Conclusion
The architecture never suffered from too many ideas; it suffered from fragments that had drifted out of relationship with one another. The useful result is therefore constitutional: every source of intelligence keeps its meaning, every decision has one accountable owner, only what survives across eras reaches capital, and durable participation is earned prospectively rather than assumed. The masterpiece is not a bigger model — it is a system that keeps the whole position in view and ships only what reality confirms.
Continue reading: Part VII — The Weave That Holds, the concluding paper on selective context weaving, evolving questions, and the sealed final model.