Grid Geometry: Dimensional Structure, Independence, and Semantic Alignment

Scope and Evidence Status

This whitepaper synthesizes six studies examining the geometric properties of the Icosa personality model’s 4x5 grid architecture. All evidence is synthetic: profiles were computationally generated by the Icosa engine, and the analyses characterize the behavior of the scoring algorithm, not human personality structure. The studies collectively produced 16 final-reportable findings, zero nulls, zero exploratory positives, and zero below-threshold results. Two findings carry circularity governance flags (both classified as expected formula-behavior checks) and are reported at their governed status throughout.

The evidence taxonomy for this synthesis is as follows. Five of the six studies are formula verification studies that check whether the engine’s output behaves as its mathematical specification predicts. One study is a synthetic benchmark that tests whether the d40 assessment instrument recovers geometric structure beyond marginal summaries. None of the studies involve human respondents, external criterion measures, or clinical outcome data. Claims about “personality structure” in this synthesis refer exclusively to the structure of the computational model’s output space. Whether that structure transfers to human populations is the central open question that this evidence base cannot answer.

The Studies

The six studies, their evidence types, and their aggregate signal profiles are:

All studies used N = 1,000 synthetic profiles generated via computational simulation (seed = 42), except the d40 benchmark which used N = 640. Holm-Bonferroni correction was applied within hypothesis families across all studies. All reportable findings survived both Holm-Bonferroni and program-level FDR correction (436 total tests across the research program).

Major Pattern 1: The Grid Is High-Dimensional but Not Orthogonal

The most consequential finding across this evidence base is a productive tension: the 4x5 grid preserves near-maximal dimensionality while its axes produce correlated summary statistics. These two properties are not contradictory, but their coexistence defines the geometric character of the model.

Dimensionality. Principal component analysis of 1,000 synthetic profiles (Grid Architecture study) yielded 18 effective dimensions for 95% cumulative variance, exceeding the 15-dimension adequacy threshold by a margin of 3. The first principal component captured 52.5% of variance, with the remaining 17 required components each contributing between 2.0% and 3.3%. This is a strong result for a 20-variable system: only 2 of 20 possible components were absorbed before the variance criterion was reached. The grid is not a decorated wrapper around a four-factor (one per capacity row) or five-factor (one per domain column) structure. Its 20 centers function as 18 distinguishable measurement channels under random input conditions.

Within-row redundancy. A representative within-row center pair (Sensitivity and Empathy, both in the Open row) correlated at r(998) = .47, p < .001, well below the .70 redundancy ceiling. Sharing a capacity row introduces meaningful covariance (approximately 22% shared variance) without collapsing adjacent centers into a single signal. The remaining 78% unique variance is what makes center-specific constructs like Traps and Gateways computationally viable at the individual-cell level.

Cross-axis coupling. Despite the high dimensionality at the center level, aggregate summary statistics show substantial inter-axis coupling. The Capacity Independence study found correlations between capacity row means (Open-Focus: r = .48) and capacity row variances (Open-Bond: r = .61), with Domain Health scores coupling at r = .54-.57. The Domain Organization study tested related but distinct metrics — Physical-Emotional Domain Health at r = .57 and Open-Focus Capacity Health at r = .61 — where the Capacity Independence study measured capacity row means while the Domain Organization study measured Capacity Health scores; the similar magnitudes are consistent but the two studies are not measuring the same variables. Effect sizes ranged from medium to large, and all exceeded the |r| = .10 practical significance floor by wide margins. Critically, the studies tested these correlations against a priori independence thresholds of |r| < .50, and three of four pairings in the Capacity Independence study and both pairings in the Domain Organization study exceeded that threshold.

Interpretation. The grid’s cells are non-redundant, but its row-level and column-level summaries share 23-37% of their variance. The scoring pipeline’s nonlinear penalty function, applied uniformly across domains within each capacity, is the most parsimonious explanation. When a profile is globally elevated or depressed in one capacity, the generation process that produces that pattern is statistically likely to produce similar patterns in other capacities. Capacity row variances show stronger coupling (r = .61) than capacity row means (r = .48) because domain conditions act as a latent common cause of within-row variance across all rows. Domain Health scores inherit this coupling bidirectionally: correlated rows inflate column aggregates, and shared column conditions inflate row aggregates.

This shared variance is a structural property of the scoring algorithm, not evidence that the axes measure the same psychological construct. The Icosa model was not designed to produce orthogonal axes; its architecture deliberately encodes cross-axis linkages through Gateways, Traps, and Basins that span both rows and columns. The Body Gate (Open x Physical) serves as the escape route for 12 Traps originating across multiple rows. Four of nine Gateways sit at intersections involving the Open and Focus rows, while Bond and Move account for the remaining five, creating cross-row wiring that plausibly contributes to their r = .61 coupling. The question is not whether axes correlate, but whether the magnitude of correlation is consistent with what the model’s structural rules predict.

Architectural implication. Any downstream computation that treats row-level and column-level health indices as independent inputs will overstate the model’s informational dimensionality. Formation classification, which uses row and column variance patterns, may conflate profiles that differ in their 20-center detail but share similar summary statistics. Coherence computation, which operates at the individual center level, avoids this problem by design.

Major Pattern 2: Column Completion Outperforms Row Completion as a Coherence Predictor

The Polarity Structure study uncovered a non-uniform relationship between structural balance and global integration. Physical Domain column completion correlated with Coherence at r(998) = .60, p < .001, while Open Capacity row completion correlated at r(998) = .51, p < .001. Both effects are large, but the .09-point gap between them is consistent across the synthetic parameter space and points to a structural asymmetry in how the Coherence formula responds to balance at different levels of the grid.

The Physical column’s outsized influence is traceable to its Gateway density: it contains both the Body Gate (Open x Physical) and the Vitality Gate (Move x Physical), two of nine Gateways that constrain system-wide integration. When these Gateways function within their effective ranges, they unblock escape routes from Traps, cascading into reduced Trap activation and higher global Coherence. A Physical column with balanced expression across all four capacities necessarily means both Gateways are operating effectively.

Corner ratio (the proportion of centers at scoring extremes) showed a medium negative association with grid completion, r(998) = -.33, p < .001. Profiles dominated by extreme scores are structurally sparse: they fail to populate the middle regions of the 4x5 matrix, leaving large areas of the personality structure undifferentiated. The effect is consistent but not deterministic, as extreme-scoring profiles can still achieve moderate grid completion if their extremes are distributed across different regions rather than concentrated.

Architectural implication. Centering Path algorithms may benefit from prioritizing domain completion, particularly in Gateway-dense columns, as a candidate optimization target. The row-column asymmetry provides a principled basis for weighting domain-level balance more heavily than capacity-level balance when selecting intervention steps. However, this finding is based on a single row and single column exemplar; comprehensive analysis across all four rows and all five columns is needed to determine whether the asymmetry generalizes.

Major Pattern 3: Geometric Position Carries Interpretive Weight

The Semantic-Geometric Alignment study tested whether the Icosa model’s hot-core versus cool-periphery topological distinction captures meaningful structural differences.

Governed result (formula-behavior check). Hot-core health correlated with Coherence at r(998) = .94, p < .001. This analysis is flagged under circularity governance: hot-core health shares computational ancestry with the Coherence formula, as eight of the core’s center health scores feed directly into the Coherence computation. The correlation confirms expected formula behavior. It should not be read as independent evidence that core centers are more important than periphery centers. The gap between r = .94 and r = 1.00 quantifies the degree to which periphery centers modulate Coherence independently of the core: approximately 11% of Coherence variance is not captured by core health alone.

Independent result. Hot-core and cool-periphery health correlated within profiles at r(998) = .89, p < .001. This coordination is not prescribed by any equation in the model. It arises from the interaction of the penalty architecture, the capacity-row structure, and the profile generator’s distributional properties. Nothing in the Coherence formula requires that core and periphery health move together; the model could in principle produce profiles where one zone thrives while the other languishes. The observed coupling indicates that the geometric partition captures a real structural feature of the model’s output space.

Equivalence result. A paired TOST equivalence check supported practical equivalence of mean health levels across the two zones within the pre-specified +/-2.0 mean-difference bound. The observed mean difference was 0.06, with the 90% equivalence interval fully inside the bound, and the corresponding point-null test was non-significant (p = .833; negligible d = 0.007). The result supports practical equivalence rather than a difference: profiles do not systematically show higher health in one zone than the other.

Architectural implication. Centering Paths can reasonably assume cross-zone coordination when sequencing intervention steps. Targeting a core Gateway may produce ripple effects that improve periphery centers. However, the 11% of Coherence variance independent of core health confirms that periphery-targeted steps remain necessary for profiles where core health is high but overall Coherence remains suppressed.

Major Pattern 4: Cell-Level Geometry Carries Non-Redundant Information Beyond Marginals

The d40 Geometry Beyond Marginals study directly tested whether the full 20-center geometry encodes recoverable structure that row and column marginals discard. Across 640 synthetic profiles constructed to share identical capacity means and domain means (differing only in residual geometric family membership), full-geometry classification outperformed marginal-only classification: McNemar’s chi-squared = 28.16, p < .001, d_b = 0.290 (small but practically significant, exceeding the 0.10 threshold).

The effect size deserves careful framing. At d_b = 0.290, the advantage is small. This is expected by design: all profiles shared identical row and column templates, so the residual signal was, by construction, the leftover after the dominant marginal structure was removed. A small effect in this regime indicates that cell-level residuals carry detectable but not overwhelming structure. Marginals do most of the work; cell interactions provide supplementary geometric discrimination.

The practical question is whether this supplementary signal matters for downstream constructs. Gateways, whose states are determined by single cells, are maximally sensitive to cell-level information by definition. Traps like Rumination (Focus x Mental) or Emotional Flooding (Bond x Emotional) depend on individual Harmony states, not general capacity-row elevation. Basins require coordinated deviations across specific cells that may leave row and column marginals unremarkable. If these constructs are to function as intended, the assessment instrument must preserve cell-level resolution. The benchmark confirms that the ICOSA-D40’s two-item-per-Harmony format preserves enough resolution for geometric family recovery.

Architectural implication. Any pipeline that reduces 20-center profiles to nine marginal values before computing Traps, Basins, or Gateways is operating on a lossy representation. Marginal compression is acceptable when cell-level constructs are not the analytic target, but inadvisable when they are.

Governed Findings and Circularity Constraints

Two analyses across the six studies carry circularity governance flags, both classified as expected formula-behavior checks:

1. Grid Architecture H3: Sensitivity health correlated with Coherence at r(998) = .72, p < .001. Because Coherence is computed as a weighted aggregate of all 20 center health scores, this correlation involves shared computational ancestry. The result confirms that the aggregation formula operates as specified; it does not constitute independent evidence of grid validity.

2. Semantic-Geometric Alignment H1: Hot-core health correlated with Coherence at r(998) = .94, p < .001. Hot-core health shares eight cell health scores with the Coherence computation. The result is an implementation-fidelity verification.

Neither governed finding has been upgraded to independent evidence status in this synthesis. They are reported as confirmations that the engine’s formulas behave as their mathematics specify.

Null and Below-Threshold Results

No null results, exploratory positives, or below-threshold findings were observed across any of the six studies. All 16 tested hypotheses moved in their predicted directions. All effect sizes exceeded their respective practical significance thresholds after correction.

The absence of nulls is itself informative: the Coherence computation behaves as the geometric logic predicts across the synthetic parameter space, without anomalous reversals, floor effects, or edge cases that collapse dimensionality. However, the uniformly positive results also reflect the evidence type. Formula verification studies test whether an engine does what its code says it does; they are designed to confirm, not to falsify. The absence of nulls should be read as implementation fidelity, not as evidence of construct validity.

The closest result to a null was the Open-Focus capacity mean correlation at r = .48, which fell just below the .50 independence threshold specified in the Capacity Independence study’s hypothesis. The remaining three correlations in that study exceeded the threshold. Whether r = .48 versus r = .50 represents a meaningful boundary or an arbitrary cutpoint is a matter of interpretation; the substantive pattern of moderate-to-large cross-axis coupling is consistent across all tested pairings.

Limitations of the Evidence Base

Four structural limitations bound the claims that can be made from this evidence:

Synthetic data only. All profiles were computationally generated. The dimensionality, coupling magnitudes, and structural relationships documented here characterize the scoring algorithm’s behavior, not human personality organization. Whether these properties transfer to human-collected data is the central unanswered question.

Random generation only. All studies used uniform random sampling (seed = 42) without persona constraints. Clinical-like profiles, which impose systematic covariance patterns, could amplify, attenuate, or redirect the baseline coupling observed here. The studies explicitly identify persona-constrained replication as a necessary next step.

Sampled axis pairs. The cross-axis coupling studies tested selected pairings (Open-Focus, Open-Bond, Physical-Emotional, Physical-Spiritual) rather than exhaustive matrices. The full 6-pair capacity correlation matrix and 10-pair domain correlation matrix remain uncomputed. Adjacent pairs were chosen where coupling was expected to be highest; non-adjacent pairs may show weaker associations.

Linear analysis only. All correlations are Pearson product-moment (linear). Nonlinear dependencies between axes or zones would not be detected. Several variables showed non-normal distributions (capacity variances had skewness exceeding 2.0), though sample sizes provided adequate stability for the linear estimates.

Architectural Implications

The synthesis identifies four engineering consequences for the Icosa framework:

1. Row and column summaries are not independent inputs. Any downstream computation that multiplies or weights Capacity Health and Domain Health independently risks double-counting shared variance. Coherence computation avoids this by operating at the center level. Narrative systems that present axis-level summaries as separate story threads should either partial out the shared component or replace axis-level descriptions with center-level ones.

2. Domain completion is a higher-leverage optimization target than capacity completion. The Physical column’s r = .60 with Coherence versus Open row’s r = .51 provides a principled basis for weighting domain balance in Centering Path algorithms, contingent on confirmation across the full set of rows and columns.

3. Cell-level resolution must be preserved for derived constructs. Marginal compression entails measurable information loss (d_b = 0.290). Traps, Basins, and Gateways require the full 20-center representation.

4. Effective dimensionality serves as a regression test. The 18-dimension baseline should be monitored during engine development. Any modification that reduces effective dimensionality below 15 warrants investigation.

Next-Step Research Priorities

The studies converge on five specific follow-up analyses, ranked by their expected yield:

1. Full pairwise correlation matrices (all 6 capacity pairs, all 10 domain pairs), stratified by Coherence Band. This would determine whether cross-axis coupling is uniform or tightens under distress, and whether the Physical column’s Coherence-predictive advantage extends to other Gateway-rich regions (e.g., the Mental column containing the Choice Gate and Identity Gate).

2. Persona-constrained replication of all six studies. Clinical-like profiles impose structured covariance patterns absent from random generation. Whether the grid’s 18-dimensional structure compresses under systematic pathology, and whether cross-axis coupling amplifies in Burdened/Severe profiles, are the most important open questions for model validity.

3. Construct-conditional decomposition of the d40 full-geometry advantage. Testing whether the marginal-vs-full accuracy gap concentrates in profiles with active Traps, Basins, or constrained Gateways versus structurally simple Thriving profiles would clarify when marginal compression is safe.

4. Perturbation analysis of the core-periphery coordination. Holding core centers fixed while varying periphery scores (and vice versa) would reveal whether the r = .89 coordination reflects bidirectional influence or unidirectional drive, with direct implications for Centering Path step sequencing.

5. Human-data benchmarking. Comparing the synthetic dimensionality (18 components), cross-axis coupling magnitudes (r = .48-.61), and core-periphery coordination (r = .89) against their human-data analogs is the definitive test of whether the model’s geometric properties describe personality structure or only describe the model’s own computational behavior.

Until priorities 1-4 are addressed within the synthetic evidence framework, the current findings establish that the Icosa engine’s geometry operates as its mathematics specify: high-dimensional, non-redundant at the cell level, coupled at the aggregate level, and structurally responsive to polarity balance. These are necessary conditions for the model’s higher-order constructs to function as designed. They are not sufficient conditions for those constructs to capture real psychological phenomena. The bridge from computational architecture to human validity requires external data that this evidence base does not contain.