Abstract
The earlier papers built a philosophy, a falsification discipline, a judgment layer, and an opponent model. In live conditions, the system still bled. Historical label discrimination remained strong, but serving-parity and economic tests did not exonerate the model or prove live profitability. The failure was partly interpretive: epistemic information, whose job is to say whether a prediction should be trusted, was allowed to answer the directional question of what the move should be.
Paper V reports the narrower conclusion that survived correction. The primary ranker retained historical ordering information. One independent wisdom signal improved the target-horizon ordering. Fixed context gates and a trap veto failed their required splits. None of these findings, alone, establishes a profitable live system.
1. The Predictor Is Conditional, Not Uniform
The ranker was reliable in some states and close to noise in others. That conditionality mattered, but no tested context partition converted it into durable matched-exposure policy lift. A model can contain information without granting its surrounding system permission to act everywhere.
2. Context Is a Lens, Not Demonstrated Alpha
Multiplying the predictor by reliability context degraded its ordering. Using that context as a gate preserved its semantic role, but the fixed gates still discarded profitable baseline trades and failed walk-forward falsification. The claim is empirical and deliberately narrow: context described when to look, but no tested intervention earned the right to control the book.
3. Captured Signals Are Directional Tells
Configurations that looked bullish at the end of a move were associated with subsequent weakness in the discovery window; frightening capitulation-like configurations were associated with strength. Our long-biased predictor bought some of the manufactured tops that the trap detectors had identified. The diagnosis was real. The permanent fade rule was not.
4. What Survived
- Rank: preserve historical ordering as a baseline, but require prospective economic validation before calling it tradeable.
- Lift: one physically grounded modifier improved the ordering at the aligned 10-day horizon, not universally at 5 or 21 days.
- Gate: context may describe reliability and participation, but the fixed gates tested here are not production policy.
- Protect: a dedicated safety judgment may reduce the left tail without pretending to add directional edge.
5. Direction Is an Authority
The final audit found multiple downstream components that could recreate, reverse, or enlarge a trade after the model had spoken. The repair is constitutional rather than predictive: the canonical model owns direction; context can delay or abstain; timing cannot change the thesis; sizing can only reduce authorized exposure; risk remains final; delayed execution must renew authorization.
6. What This Paper Deliberately Does Not Disclose
We withhold the confirmations, wisdom signal, trap composition, safety mechanics, coefficients, data universe, and per-experiment magnitudes. The paper shares the failure boundary and the authority rules without handing over the instruments that would enable direct reproduction.
Conclusion
Prediction is the output of intelligence; judgment is the beginning of wisdom; execution is where both can be quietly contradicted. Removing category errors is necessary, but it is not alpha. The honest result is smaller: preserve the rank, align the lift to its horizon, keep epistemic information out of directional computation, and give one canonical model ownership of direction.