Construct Dynamics: Traps, Basins, Gateways, and Fault Lines

Scope and Evidence Status

This whitepaper synthesizes findings from seven studies that benchmark the construct system of the Icosa personality model under synthetic conditions. All data were generated computationally using the Icosa engine’s d40 assessment pathway, with sample sizes of N = 167 (one study) and N = 1,000 (six studies). No human participants were assessed. The evidence base consists of formula verification studies (six of seven) and one synthetic benchmark study. All circularity audits returned clean: zero flagged, zero unresolved, and zero disallowed findings across the entire program. These results therefore speak to the internal consistency of the engine’s computational logic — whether the constructs behave as the geometry specifies — and do not constitute evidence of external validity, diagnostic performance, or treatment effectiveness.

The evidence taxonomy governs interpretation throughout. Formula verification studies confirm that the engine implements its own rules faithfully. The synthetic benchmark (equifinal paths) tests a structural hypothesis within the model’s simulated parameter space. Neither category supports claims about human personality functioning. Where findings are characterized as “large” or “strong,” these descriptors refer to effect magnitudes within synthetic distributions, not to clinical or population-level importance.

Across the seven studies, 65 hypothesis tests were conducted. Of these, 51 reached final reportable significance after Holm-Bonferroni correction and met pre-registered practical significance thresholds. Three were statistically significant but fell below practical thresholds. One reached nominal significance only (raw-only). Ten were null. Zero were classified as exploratory positive or not evaluable. This yield — 78% fully reportable, 15% null, 6% mixed or sub-threshold — establishes a broad but bounded verification of the construct system.

The Construct System Under Test

The Icosa model derives four families of higher-order constructs from a 4x5 grid of 20 Harmony centers (four Capacities crossed with five Domains):

• Traps (50 defined): Self-reinforcing feedback loops at individual centers, activated when a center deviates sufficiently from its capacity-specific target while its designated escape gateway remains closed.
• Basins (32 defined): Multi-center attractor states that resist perturbation, requiring coordinated dysfunction across four or five centers.
• Gateways (9 channels): Structurally critical intersections of Capacity rows and Domain columns that regulate system-wide dynamics. Each gateway can be open, closed, or in an overwhelmed or paradoxical state.
• Fault Lines: Cascade-vulnerability patterns defined by displacement across specific center subsets, predicting structural fragility.

These constructs are not independent parameters but emergent consequences of grid geometry. The studies collectively test whether this emergence is lawful — whether the constructs co-vary, separate, and interact as the underlying mathematics requires.

Trap Emergence and Taxonomy

Two studies examined the origins and categorical structure of traps.

Geometric origins. The trap-emergence study verified that trap activation follows from grid-level variance rather than absolute score levels. Variance penalty correlated with Trap Count at r_s(998) = .57, p < .001, a large effect confirming that the engine’s trap-detection rules are geometrically grounded. All five Domain Health indices showed large negative correlations with Trap Count, ranging from r(998) = -.59 (Physical, Emotional) to -.62 (Mental). The narrow spread of .030 units across all five domains is a structurally consequential finding: trap formation is a whole-grid phenomenon, not a domain-localized one. No single domain disproportionately drives trap emergence, though the Mental Domain’s slight numerical edge aligns with the Choice Gate’s outsized load (15 of 50 trap escape routes). Shadow and oscillation constructs co-occurred at medium effect, r_s(998) = .38, confirming shared geometric ancestry without redundancy.

Taxonomic separation. The trap-taxonomy study tested whether the model’s five domain assignments (Physical, Emotional, Mental, Relational, Spiritual) encode genuine computational distinctions. All 11 hypotheses were confirmed. Somatic Freeze and Emotional Flooding produced reliably distinct severity distributions (mean difference = 0.11, p < .001, delta = 0.107). Identity Dissolution severity predicted centered-count depletion at r_s(998) = -.55, a large effect indicating that identity-category traps carry distinctive weight in system integration. Domain Health correlations replicated the emergence study’s pattern exactly (r range: -.59 to -.62), providing cross-study consistency within the synthetic benchmark program.

Cell-level effects at four structurally critical grid positions were smaller but reliable (r range: -.16 to -.23). The two Mental-column cells (Focus x Mental and Open x Mental) produced the strongest cell-level associations (r = -.23 each), while the Bond x Emotional cell produced the weakest (r = -.16), consistent with the Mental Domain’s outsized gateway load. The gap between domain-level effects (large, r ~ -.60) and cell-level effects (small, r ~ -.20) is itself informative. If trap activation were strictly local — driven by the originating cell alone — cell-level and domain-level correlations should be comparable. The three-fold difference indicates that traps are activated by distributed patterns across multiple cells within a domain, not by single-cell deviations. This distributed activation pattern is consistent with the network propagation model underlying the Icosa architecture, where bridge centrality between domains enables dysfunction to spread beyond its point of origin.

The gap also raises a precise and testable question about the Feeling Gate (Bond x Emotional), whose cell showed the weakest association with Trap Count despite the Emotional Domain being one of the most gateway-dense domains. Whether this reflects the Feeling Gate’s influence operating as a continuous gradient or a discrete threshold is resolvable through boundary-condition simulation.

Replication note. The Domain Health correlations with Trap Count are identical between the trap-emergence and trap-taxonomy studies because both used the same N = 1,000 synthetic sample (seed = 42). These are shared-anchor results, not independent replications. The agreement confirms computational reproducibility but should not be counted as converging evidence from separate data sources.

Basin Discovery and Domain Independence

The basin-discovery study tested the relationship between basins and traps and the impact of basins on Coherence (the model’s 0-100 integration metric).

The central finding was the inverse association between Basin Count and Coherence, r_s(998) = -.51, p < .001. This large effect confirms that the engine’s Coherence formula is structurally sensitive to multi-center attractor states. Basin Count and Trap Count were positively correlated at r_s(998) = .22, a small effect. The disparity between the basin-Coherence effect (-.51) and the basin-trap effect (.22) is informative: basins degrade Coherence through channels beyond simply co-occurring with traps. The structural inertia that basins represent — multiple centers mutually reinforcing dysfunctional states — imposes integration costs captured through the Coherence formula’s cross-center interaction terms.

Null results. Six of nine hypotheses were null, and these nulls are substantively important. No individual Domain Health score predicted Basin Count (all |r_s| < .04, ps > .22). This domain-independence finding means that basins emerge from cross-domain coordination failures rather than from degradation within any single Domain column. Basin Count also did not predict hidden masked fragility beyond what grid completion already provides (delta = -.023, p = .789), establishing that basins are not a latent fragility indicator but a configurational construct operating at a different level of analysis than global severity. The only domain-specific effect to survive correction was a small negative association between Relational Domain Health and Trap Count (r_s = -.15), while Emotional and Spiritual domains showed no association with trap susceptibility.

Gateway Architecture

Three studies examined gateway behavior from complementary angles: dimensional taxonomy, downstream consequences, and interaction with other constructs.

Cascade Fidelity Across All Nine Channels

The gateway-taxonomy study tested whether open gateway states reliably predict cascade scores. All nine channels passed. Effect sizes were uniformly large, ranging from d = 1.01 (Choice Gate, Focus x Mental) to d = 1.15 (Feeling Gate, Bond x Emotional). The narrowness of this range — 0.14 d units — is the most structurally significant aspect of the result. The Choice Gate serves as the escape route for 15 traps; the Voice Gate serves zero. Yet the Voice Gate (d = 1.12) exceeded the Choice Gate (d = 1.01) in cascade magnitude. Cascade propagation is therefore computationally independent of the trap-escape mapping. The engine implements two distinct geometric mechanisms under the gateway construct: cascade (a dynamic, state-dependent propagation from a center to its structural neighbors) and trap escape (a static structural feature specifying which feedback loops a gateway can break). Both derive from grid geometry but index different properties of the system.

Impairment-Trap and Cascade Pathways

The gateway-outcome study tested downstream consequences across 19 hypothesis comparisons — the largest single study in the program. Gateway impairment predicted Trap Count for eight of nine gateways after correction, with effect sizes in the small range (r_s = .13-.19). The Body Gate (Open x Physical) and Grace Gate (Open x Spiritual) both showed r_s = .16; the Discernment Gate (Focus x Emotional) followed at r_s = .15. The Vitality Gate (Move x Physical) produced the strongest impairment-trap effect (r_s = .19) despite serving only one trap escape route, suggesting it functions as a sentinel of broad grid dysfunction rather than operating through its local escape-route mechanism. Small individual effect sizes are expected here: each gateway mediates only a subset of the 50 total traps, so a bivariate association between one gateway and the total count is necessarily diluted by traps linked to other gateways.

The cascade pathway showed a markedly different pattern, and this asymmetry is the study’s most theoretically consequential finding. Only four gateways reached practical significance for predicting dynamics cascade: the Discernment Gate (r_s = -.12, the largest), the Choice Gate (r_s = -.10), and two others at threshold. The remaining five, including the two Move-row gateways (Vitality and Voice) that showed strong impairment-trap effects, produced null or negligible cascade results. At the Body Gate, the cascade association (r_s = -.062, p = .048) reached nominal significance but did not survive correction — a raw-only result. At the Grace Gate, the cascade result was fully null (r_s = -.059, p = .062).

This divergence between the impairment pathway (8 of 9 gateways verified) and the cascade pathway (4 of 9 at practical significance) identifies a concrete boundary in the current implementation. The engine’s local bottleneck logic — gateway blocks trap escape, impairment at the bottleneck correlates with trap persistence — is verified. Its global propagation logic — gateway opening triggers system-wide cascade dynamics — is partially present but involves mediating pathways that bivariate association cannot capture. System-level cascade behavior likely depends on the configuration of open gateways (which combinations are open simultaneously) rather than any single gateway’s state.

Aggregate gateway impairment captured hidden destabilization risk beyond grid completion (delta = .09, p < .001), confirming that the gateway construct adds structural information not redundant with global severity.

Signal profile detail. Of the 19 tests, 13 were fully reportable, 3 were statistically significant but below practical thresholds (including Choice Gate impairment at r_s = .088, just below the .10 threshold), 1 reached raw significance only (Body Gate cascade, r_s = -.062, p = .048), and 2 were null (Grace Gate cascade and Vitality Gate cascade). The below-threshold and raw-only findings are preserved here because they mark the boundary between the engine’s local bottleneck logic (verified) and its global propagation logic (partially verified, partially null).

Gateway Buffering

The construct-interaction study tested gateway buffering directly. Profiles with an open Choice Gate carried dramatically fewer traps than those without, t(788.24) = 27.87, p < .001, d = 1.72. Comparable effects emerged for the Feeling Gate (d = 1.60) and the Body Gate (d = 1.67). The consistency of effect sizes across channels spanning different Capacities and Domains (Focus x Mental, Bond x Emotional, Open x Physical) suggests that the gateway buffering mechanism operates with comparable strength regardless of which capacity-domain intersection is involved. This uniformity is not trivially expected, and it provides a concrete engineering constraint: the Centering Path algorithm can treat gateway opening as a uniformly effective structural intervention.

Construct Interactions

The construct-interaction study provides the most comprehensive single test of the system’s internal consistency. All six hypotheses were confirmed with large effects:

• Fault Lines and Basin Count: r_s(998) = .85, p < .001
• Gateway bonus and Trap Count: r_s(998) = -.86, p < .001
• Fault Line count and centered count: r_s(998) = -.90, p < .001
• Gateway buffering across three channels: d = 1.60-1.72

The Fault Line-centered count association (r_s = -.90) was the largest single effect in the entire program. Each additional Fault Line corresponds to a measurable reduction in the grid’s overall centered capacity, consistent with the dynamic-systems view that cascading vulnerabilities erode system-wide stability. The near-linear depletion pattern confirms that the Fault Line detection algorithm captures something functionally consequential within the model architecture.

The strength of construct co-variation (all r_s ≥ .85 or all d > 1.60) indicates that traps, basins, gateways, and Fault Lines form an integrated geometric architecture rather than a set of loosely related indices. Dysfunction-related constructs tightly co-vary, and gateways operate as predicted structural regulators.

Equifinal Path Discrimination

The equifinal-paths study — the only synthetic benchmark rather than formula verification — asked whether different activation routes into the same trap produce distinguishable structural neighborhoods.

The answer is partially yes. Each trap admitted, on average, nearly three structurally distinct non-overlap neighborhoods above baseline (d = 1.276, p < .001), a large effect establishing within-trap heterogeneity. Path index and activation confidence predicted neighborhood signature at R^2 = .072, a small but reliable effect surviving correction.

Null results. Two of four hypotheses were null, and both matter. Perturbation response did not differ by path (mean difference = 0.01, p = .686, delta = 0.011), meaning that while equifinal paths produce structurally distinguishable configurations, those configurations do not respond differently to simulated centering interventions within the current model. Path-based predictors also did not generalize to independently generated masked fragility (R^2 = .009, p = .670). These nulls bound the functional reach of path discrimination: it is a local structural feature, not a system-level vulnerability indicator. Two profiles sharing the same trap label can occupy meaningfully different structural contexts, but those contexts do not (yet) predict differential response to intervention or independently computed risk.

This study used a smaller sample (N = 167 vs. 1,000) and a different evidence type (synthetic benchmark vs. formula verification), making its positive findings the most tentative in the program while also the most architecturally novel.

Null and Sub-Threshold Findings: Summary

Across the program, 10 null findings, 3 below-practical-threshold findings, and 1 raw-only finding deserve consolidated attention:

Domain-to-basin pathways (3 nulls). No individual Domain Health score predicted Basin Count. Basin formation is domain-independent within the engine.

Domain-to-trap pathways (2 nulls). Emotional and Spiritual Domain Health did not predict Trap Count in the basin-discovery study’s domain-specific analysis, though they did in the trap-emergence study’s whole-grid analysis. This discrepancy likely reflects differences in what each study controlled for, but it limits the claim that every domain independently contributes to trap susceptibility.

Hidden fragility (2 nulls). Neither Basin Count nor path-based predictors generalized to independently generated masked fragility. Hidden fragility appears to operate through mechanisms not captured by the constructs tested.

Perturbation response (1 null). Equifinal paths do not produce differential perturbation response. Path discrimination is structural, not functional, within the current perturbation model.

Cascade pathway (5 sub-threshold or null). Five of nine gateways failed to reach practical significance for predicting dynamics cascade. The cascade mechanism appears to depend on gateway configurations rather than individual gateway states.

These nulls are not failures of the construct system. They define its boundaries: where the geometry predicts and where it does not, where bivariate association captures the mechanism and where configural modeling is needed.

Architectural Implications

The synthesis supports four architectural conclusions within the synthetic evidence base:

1. The construct system is internally consistent. Across 51 fully reportable findings and seven studies, the four construct families (traps, basins, gateways, Fault Lines) co-vary as the geometry specifies. No construct behaves in a way that contradicts its geometric derivation. This is a necessary condition for proceeding to external validation but is not sufficient evidence that the constructs capture real personality phenomena.

2. Trap formation is a whole-grid phenomenon. The near-uniform Domain Health correlations (-.59 to -.62 across five domains) and the domain-independence of basin formation together establish that dysfunction distributes across the grid rather than concentrating in any single domain. The Centering Path algorithm should favor multi-domain strategies over single-column approaches.

3. Gateways implement two separable mechanisms. Cascade propagation and trap-escape mapping are computationally independent, as demonstrated by the Voice Gate’s strong cascade fidelity (d = 1.12) despite having zero trap escape routes. Local bottleneck logic (gateway impairment predicts nearby trap counts) is verified; global propagation logic (gateway opening triggers system-wide cascade) is partially verified and needs configural extension.

4. Basins are the primary drivers of Coherence degradation. The r_s = -.51 basin-Coherence association, combined with the modest basin-trap correlation (r_s = .22), positions basins as the construct family with the most direct impact on the model’s integration metric. Basin disruption should take precedence over isolated trap resolution in intervention sequencing.

Limitations

All findings are bounded by the synthetic data paradigm. The engine’s internal consistency does not guarantee that the constructs correspond to measurable states in human personality. Effect sizes reflect the behavior of computationally generated profiles within a single model architecture, and they may not transfer to empirical data where measurement error, response bias, and between-person variation introduce noise that the synthetic benchmark cannot simulate.

The six formula verification studies test whether the engine implements its own rules. They cannot discover whether those rules are good rules. A formula can be implemented perfectly and still fail to capture the phenomenon it was designed to model. External validation — comparing Icosa construct states to established personality measures, clinical assessments, or behavioral outcomes — is the necessary next step and the only way to move these findings from engineering verification to scientific evidence.

The shared N = 1,000 sample (seed = 42) across six studies creates correlated error structures. Findings that appear to converge across studies may partly reflect shared sampling characteristics rather than independent confirmation. The equifinal-paths study used a separate sample (N = 167) and is the only study with an independent data source within this program.

Next-Step Research Priorities

1. External validation. The highest priority is testing whether construct states derived from human assessment data reproduce the patterns observed in synthetic benchmarks. Specifically: does the basin-Coherence association survive in human profiles, do gateway states predict trap counts in empirical data, and do domain-health correlations with trap formation retain their near-uniform character when measurement error is present?

2. Configural gateway analysis. The bivariate cascade findings (4 of 9 gateways at practical significance) demand a configural follow-up: which combinations of open gateways predict system-wide dynamics cascade? This is the most direct route to closing the gap between verified local bottleneck logic and partially verified global propagation logic.

3. Basin perturbation study. The aggregate basin-Coherence association (r_s = -.51) should be decomposed by computationally dissolving individual basins and measuring per-basin Coherence impact. This would determine whether Centering Path basin-disruption priorities should be uniform or weighted by basin identity.

4. Boundary behavior testing. The tight coupling across construct families (r_s ≥ .85 for Fault Lines-basins, r_s = -.90 for Fault Lines-centered count) raises a stability question: do these relationships degrade gracefully at parameter-space boundaries or exhibit sharp discontinuities? This determines whether the Centering Path algorithm can rely on construct interactions as stable optimization targets.

5. Path discrimination enrichment. The equifinal-paths study’s positive findings (within-trap heterogeneity) paired with its nulls (no differential perturbation response) suggest that replacing the ordinal path index with a dimensional encoding of surrounding center states could recover the functional reach that the current encoding lacks. This is a targeted model improvement, not a validation study.

6. Multivariate domain decomposition. The uniform bivariate domain-trap correlations may mask structurally richer mechanisms. If domains contribute unique variance through their respective gateways, a multivariate decomposition would reveal whether the Centering Path algorithm should weight gateway-opening interventions by their domain-specific trap reduction potential.

Conclusion

Seven synthetic-evidence studies, comprising 65 hypothesis tests across 51 fully reportable, 4 sub-threshold, and 10 null findings, establish that the Icosa model’s four-construct architecture operates as an internally consistent geometric system. Traps emerge from variance, not absolute levels. Basins drive Coherence degradation beyond what traps alone predict. Gateways implement separable cascade and escape mechanisms with uniform strength across all nine channels. Fault Lines predict centered-count depletion with the largest effect in the program. The null findings define boundaries with equal precision: basins are domain-independent, equifinal path discrimination is structural but not functional, and cascade propagation depends on gateway configurations that bivariate analysis cannot capture. These are engineering verification results. They confirm that the computational engine implements its theoretical specifications faithfully. Whether those specifications capture something real about human personality remains an open question that only external validation can answer.