Abstract
Papers I-VI established a research program built around adversarial capture, pre-specified falsification, independent judgment, opponent interpretation, authority separation, and semantic parity. Paper VII tests the remaining architectural question: once context is preserved, how should its parts relate? We compared a shared predictive-state baseline with isolated role-specific message models, an expanded relational lattice, and an Oracle-centered semantic merge. Evaluation used chronological folds, persistence baselines, ownership-shuffled placebos, and uncertainty coverage. Trading outcomes were excluded from the state-model comparison.
The shared hybrid state model survived every pre-specified fold. The isolated-message replacement and the higher-order lattice did not. Correctly owned messages nevertheless beat ownership-shuffled messages in every receiver-fold comparison, supporting a narrower relational claim: semantic ownership carries information, but the messages do not replace the shared board. A subsequent Oracle merge was killed as redundant. Preserving the strongest opposing Truth produced a useful semantic object, but its proposed economic ranking value failed on fresh evidence. The production result is therefore constitutional, not a new alpha claim: preserve role, provenance, contradiction, and authority; leave unqualified relationships observational.
1. Research Question and Competing Hypotheses
Paper VI showed that strategic context could survive the boundary between historical replay and live inference without being flattened into unrelated columns. That solved representation parity, not composition. The capstone question was whether relations among opponent identity, directional state, epistemic trust, timing, and risk contained information beyond the preserved state itself.
- Shared-state hypothesis: one preserved board contains the relevant relations; additional cross-role messages are redundant.
- Isolated-weave hypothesis: role-owned strands can replace the shared board while preserving or improving forecast skill.
- Relational hypothesis: correctly routed messages carry information, but only as additions to a shared board.
- Expansion hypothesis: broad pairwise and higher-order interactions recover relations that sparse routing misses.
Each claim had a different way to lose. In particular, evidence for the relational hypothesis could not rescue a failed replacement model, and a state-forecast result could not be reported as trading alpha.
2. Evaluation Design
The experiments were run over ordered, out-of-sample folds. Discrete opponent and playbook states were compared with train-only transition baselines; continuous directional and epistemic states were compared with persistence. Uncertainty intervals were required to maintain coverage rather than appear precise by becoming too narrow.
To test whether semantics rather than message volume produced the result, cross-role messages were compared with deterministic within-date ownership shuffles. The sender could not pass raw foreign fields, decision outputs, or a universal latent score. Each receiving role owned its own residual interpretation. The higher-order experiment used the same underlying rows and included a linear-message comparator, so additional interaction count could not masquerade as relational value.
3. The Baseline That Survived
The first predictive-state candidate used learned identity classifiers and was killed because it could not beat simple transition rules for discrete state. The second candidate assigned discrete transitions to the simpler state process, continuous innovations to role-separated estimators, and uncertainty to a separate calibration layer. That hybrid survived every floor in every chronological fold.
This was evidence that the strategic state had causal continuity and could be forecast more accurately than persistence under the tested representation. It was not evidence that the forecast improved trades. Two separate action extensions were later tested and killed: one did not intervene on the eligible decisions, and another improved a local exit measure in one period while degrading shared-capital results. The state result survived; the economic authority claim did not.
4. The Replacement We Killed
The first weaving experiment separated identity, directional, and epistemic semantics into role-owned strands. Correctly routed messages beat ownership-shuffled controls for both receiving roles in all five folds. This result rejected the null that ownership was irrelevant.
The complete candidate still failed its primary claim. Directional and epistemic skill regressed in required folds, and uncertainty intervals widened in several folds. Semantic isolation had removed board state that the messages could not reconstruct. We therefore killed the claim that isolated strands should replace the shared hybrid state, while retaining the narrower observation that ownership-preserving communication carried relational information.
5. More Interactions Made the Weave Worse
A second pre-specified test preserved the hybrid state as the board and expanded cross-role relations into pairwise and selected higher-order products. It lost to the hybrid baseline for every receiving role in every fold and lost to the simpler linear-message comparator in every direct comparison. Its uncertainty intervals also widened throughout.
The failure identified a specific boundary: interaction scale is not evidence of relationship. Broad expansion manufactured links whether or not a physical or causal law supported them. The surviving design constraint is therefore selective routing with semantic ownership and conservation, not exhaustive combination.
6. The Oracle Was Not a Universal Weaver
Because the Oracle already owns named scenario interpretation, we next tested whether opponent and playbook transitions should be merged there. The candidate added almost no information beyond the Oracle's own state transition and worsened important subgroups. The merge was killed as redundant.
The failure exposed a different omission. The Oracle retained the winning explanation but discarded the strongest declared alternative. We preserved that nearest opposing Truth as a counter-Truth, with explicit evidence coverage and no automatic decision authority. This was an architectural result: the system could retain the proposition most capable of falsifying its current interpretation.
7. Counter-Truth Survived Semantically, Not Economically
We then tested whether the distance between winning Truth and counter-Truth added economic ordering information after controlling for the named Truth and the primary model's rank. The first fresh cohort was too small and was reported as pending rather than scored. A broader pre-specified cohort matured and failed the required chronological and symbol splits. The economic exclusivity claim was killed.
This negative result defines the discovery more precisely. Counter-Truth is retained because it conserves the unresolved alternative and supports a falsifiable next question. It is not a demonstrated ranker, gate, or sizing signal. The system may propose one question about the contested relationship, but the proposal remains inactive until a separate fresh test earns authority.
8. What Entered the Final Model
The production release incorporates the findings as constraints rather than as an unverified cohesion score. The primary model retains ranking authority. The Oracle retains interpretation and final admission authority. Timing and exit remain with their designated owner; risk capacity remains separate. Preserved context, state forecasts, counter-Truth, and evolving-question proposals cannot silently acquire those powers.
The release is sealed across its learned artifacts, semantic contracts, authority path, and serving transport. Future research remains possible, but a new observation cannot change behavior until its claim is pre-specified, tested on fresh evidence, reviewed, and bound into a new release. This is a production-governance conclusion, not evidence that every retained observation contributes alpha.
9. Findings Across the Series
- Supported so far: constraint-grounded questions and preserved semantic context provide a more stable research basis than confidence aggregation alone.
- Supported in the capstone: a hybrid predictive state beat its declared baselines, and correctly owned messages beat ownership-shuffled controls.
- Killed in the capstone: isolated strands as a replacement, dense relational expansion, a redundant Oracle state merge, and counter-Truth distance as economic ordering alpha.
- Retained as engineering obligations: train-serve parity, explicit unknown states, role ownership, authority separation, immutable failure records, and fail-closed release verification.
- Still unresolved: whether a future law-declared relation, observed prospectively at sufficient breadth, can earn incremental economic authority without weakening the shared board.
10. What This Paper Deliberately Does Not Disclose
We withhold the model artifacts, feature and question mappings, Truth predicates, counter-Truth selection rules, transition states, calibration methods, thresholds, receiver routes, promotion gates, data universe, and measured edge magnitudes. Those details determine how the weave is constructed and would enable direct reconstruction. We share the research principles, failure boundaries, and governance that others can evaluate without publishing the instruments themselves.
Conclusion
The capstone did not validate a model that combines everything. It rejected three versions of that idea: isolated replacement, exhaustive interaction, and redundant semantic merging. The result that remained was narrower. Shared state matters. Ownership-preserving relations contain information. Neither fact grants a relation decision authority.
The final model therefore closes the series as a governed research result. It preserves the board, retains the strongest unresolved contradiction, and allows that contradiction to propose a new falsifiable question. Production behavior changes only when fresh evidence survives the same rules of proof. The weave that held was not the largest one; it was the one that preserved meaning while remaining willing to be wrong.