Centering Paths: Intervention Sequencing, Overlay Effects, and Practical Translation

Scope and Evidence Status

This whitepaper synthesizes five studies benchmarking the Centering Paths subsystem of the Icosa personality model. All evidence is synthetic: profiles were computationally generated, interventions were simulated as deterministic state transitions, and outcomes were measured using model-internal metrics. No human respondent data, clinical outcomes, or treatment effectiveness claims are involved. The studies characterize the computational architecture’s internal consistency, identify boundary conditions, and establish benchmark baselines against which future empirical work can be evaluated.

The five studies span three evidence types. One simulator comparison tested whether intervention ordering affects cumulative trajectories (N = 2,000). Three formula verification studies tested whether path-level dynamics, overlay modulation channels, and path efficiency metrics behave as the model’s mathematics specify (each N = 1,000). One applied synthetic study tested whether path outputs correspond to interpretable health signals within the model’s own profile space (N = 1,000). Across all studies, 21 analyses were conducted. Nineteen produced reportable confirmatory findings after Holm-Bonferroni correction. Two were null. One analysis carried a circularity governance flag. No analyses were not-evaluable, below practical threshold, or exploratory-only.

The Centering Path Algorithm

The Icosa model computes Centering Paths as ordered intervention sequences across a 4x5 geometric personality structure (four capacity axes by five domain columns, yielding 20 centers). The algorithm follows a priority hierarchy: Gateway unlocking, then Basin disruption, then Trap escape, then direct Coherence improvement. Gateways are specific grid positions whose centered states release dependent Traps, creating cascading downstream gains. Basins are attractor states that resist change through coordinated pull across multiple centers. Traps are individual centers locked in extreme positions by structural dependencies. Coherence, scored 0-100, is the model’s primary integration metric.

The path engine generates milestone sequences, tracks compensatory adjustments when shifting one center displaces a neighbor, records oscillatory reversals when competing structural constraints force back-and-forth movement, and optionally computes dual-path configurations that offer alternative entry points into the system. The five studies benchmarked these properties under controlled synthetic conditions.

Path Dependence: Ordering Matters

The centering-plan-policy-benchmark study addressed a foundational question: does the order in which identical interventions are applied affect the cumulative Coherence trajectory, or does any permutation converge to the same outcome?

Across 2,000 synthetic profiles, gateway-ordered sequences produced significantly higher cumulative Coherence (measured as area under the curve) than randomly-ordered sequences applying the same interventions. A permutation paired test confirmed the advantage (mean difference = 0.34, p < .001, delta = 0.345), and bootstrap analysis quantified its magnitude (mean difference = 0.34, p < .001, d = 0.423, 95% CI excluding zero). The effect is small by conventional standards but exceeded the pre-registered practical threshold of 0.05 by a wide margin. Both analyses survived Holm-Bonferroni and FDR correction.

The convergence between permutation and bootstrap frameworks on the same mean advantage of 0.34 strengthens confidence in the estimate’s stability. The effect size indicates that the Icosa model’s state space is path-dependent under intervention: Gateways function as genuine structural leverage points whose early resolution amplifies the Coherence yield of subsequent steps. When the Choice Gate (Focus x Mental, escape route for 15 Traps) is unlocked early, dependent Traps that are addressed later produce larger gains than they would with the Gate still closed. Random sequencing forfeits this cascading benefit.

The practical implication within the synthetic benchmark is that the gateway-first heuristic is internally justified and not a simplification that can be dropped without cost. The d = 0.423 baseline provides a concrete reference for evaluating alternative sequencing policies (Basin-first, Fault-Line-priority, hybrid approaches) using the same AUC framework.

Null result. A secondary hypothesis tested whether gateway-ordered AUC captured latent risk reduction beyond baseline profile severity. It did not. The correlation comparison between gateway AUC and grid completion as predictors of hidden future instability yielded a negligible difference (delta |r| = -0.01, p = .646, well below the 0.03 practical threshold). Ordering improves the trajectory path but does not alter the endpoint vulnerability signature. The information gateway sequencing adds concerns the journey — fewer destabilizing intermediate states — rather than the destination’s residual risk profile.

Internal Consistency: Sequential Dynamics

The intervention-ordering study verified that two core sequential dynamics of the path algorithm — compensation and oscillation — behave consistently with the model’s mathematical commitments.

Compensation fidelity. Compensation count (discrete steps where shifting one center displaces a neighbor) correlated strongly with the dynamics layer’s continuous compensation metric: rs(998) = .64, p < .001. This confirms that two independent computations — one sequential and step-level, the other continuous and aggregated — draw on compatible geometric information. The correlation is strong but not perfect, with residual variance likely reflecting differences in how the two systems weight correction magnitude versus frequency. A path with few but large compensatory adjustments produces high dynamics compensation with low compensation count, introducing legitimate divergence.

Oscillation-Coherence coupling. Oscillation count (directional reversals in the path) showed a strong inverse association with Coherence: rs(998) = -.62, p < .001. Profiles whose paths contained more back-and-forth movement had lower overall integration. This confirms that oscillatory path signatures correspond to structural complexity — competing Basins, contradictory Gateway constraints, or Traps pulling in opposing directions — rather than algorithmic noise. A profile caught between the Anxious Gripping Basin and the Detached Surveillance Basin would generate oscillatory steps as the algorithm alternates between addressing each attractor. That such profiles also score lower on Coherence is a non-trivial result: the Coherence formula is sensitive to exactly the structural entanglement that produces oscillatory intervention sequences.

Both effects survived Holm-Bonferroni correction and exceeded their pre-registered practical significance thresholds. No hypotheses in this study were null.

Overlay Modulation Channels

The overlay-effects study tested whether the dynamics layer’s overlay metrics — cascade and compensation — maintain lawful relationships with the path-level constructs they modulate.

Cascade-trap linkage. Dynamics cascade correlated with Trap Count at rs(998) = .87, p < .001, a large effect. This analysis carries a circularity governance flag: it is classified as an expected formula-behavior check because the cascade derivation and trap detection share computational ancestry. The result confirms implementation fidelity — the cascade metric is a near-lossless summary of trap-relevant propagation structure — but it cannot be interpreted as independent empirical support. It verifies that the engine does what its formulas specify, not that the formulas themselves are valid descriptions of external phenomena.

Compensation channel (independent). Dynamics compensation correlated with compensation count at rs(998) = .64, p < .001. No circularity flag applied to this analysis: the compensation overlay metric and compensation count are computed through distinct pathways, making this an independent verification. The effect magnitude matches the identical rs = .64 found in the intervention-ordering study for the same variable pair, providing cross-study replication within the synthetic benchmark program.

The magnitude gap between the two channels is informative. Cascade’s rs of .87 approaches what a derived-from-shared-inputs metric can achieve against a related count variable. The compensation channel’s rs of .64, while large, leaves more unexplained variance — likely reflecting an asymmetry in the constructs. Compensation involves cross-center state adjustments that may be partially orthogonal to a simple count of compensatory patterns. A profile can have high compensation dynamics distributed across few discrete patterns, or accumulate many patterns through minor, low-magnitude adjustments. The overlay metric and the count variable measure related but non-identical aspects of the phenomenon.

Path Efficiency and Scaling

The path-efficiency study tested whether the algorithm generates milestone sequences whose length tracks the model’s own integration metric and whether dual-path computation expands the intervention space.

Milestone-Coherence scaling. Milestone count correlated with Coherence at rs(998) = .73, p < .001, a large effect exceeding the pre-registered threshold of rs >= .50. The Gateway-first priority hierarchy produces path lengths that are not arbitrary but track integration in a predictable, monotonic fashion. This is not trivial: a greedy algorithm maximizing local Coherence gain at each step could produce paths that correlate weakly with global Coherence if local optima conflict with global structure. At rs = .73, the priority scheme avoids this failure mode across a broad parameter space. Roughly 27% of unexplained variance likely reflects profiles where Basin inertia or Fault Line vulnerability creates nonlinear resistance that a monotonic path-length metric cannot fully represent.

Dual-path expansion. Dual-path configurations generated significantly more intervention steps than single-path configurations (mean difference = 0.59, p = .002, delta = 0.591). The additional steps represent structurally distinct intervention targets, not redundant copies of single-path milestones. Computing two complementary trajectories expands the available entry points into the system in a non-redundant way, consistent with dynamical systems perspectives emphasizing that multiple perturbation routes can reach the same attractor state.

Both results survived Holm-Bonferroni and FDR correction. The milestone-Coherence correlation provides a quantitative baseline: any modification to the priority hierarchy, Gateway dependency graph, or Basin disruption logic can be benchmarked against rs = .73 as the current fidelity floor. Degradation below the .50 practical threshold would indicate a broken load-bearing property.

Practical Translation: Health Signal Correspondence

The practical-translation study tested whether path outputs correspond to interpretable profile-level health signals, bridging the gap between raw algorithmic output and structurally meaningful input features.

Grid completion and path availability. The most structurally revealing finding across all five studies: grid completion and single path count showed a near-perfect inverse association, rs(998) = -.96, p < .001. Path availability is almost deterministically constrained by assessment coverage. Incomplete assessments do not merely reduce precision; they structurally limit what the algorithm can compute, because Gateway-dependent cascade logic requires scorable data at specific grid positions. A profile missing data at key Gateway positions cannot trigger the multi-step cascade logic.

Aggregate health tracking. Hot core health predicted centered count with a large effect, rs(998) = .59, p < .001, confirming that the algorithm’s input-output mapping preserves meaningful health information at the aggregate level.

Granular health contributions. All five Domain Health scores (Physical, Emotional, Mental, Relational, Spiritual) correlated with centered count in the expected direction, with rs values ranging from .21 to .27 (all p < .001, all surviving Holm-Bonferroni correction). All four Capacity Health scores (Open, Focus, Bond, Move) correlated with single path count at rs = -.22 to -.27 (all p < .001, all surviving correction). The effects were consistently small.

The aggregate-versus-granular asymmetry is theoretically informative. Hot core health at rs = .59 substantially exceeds any individual domain or capacity channel (none above rs = .27). This pattern is expected under a model where Coherence emerges from cross-domain, cross-capacity interactions rather than summation of independent channels. A profile could have moderately healthy scores in every domain yet have low centered count if capacity flow axes are misaligned.

The uniformity of domain-level effects (rs = .21-.27) was unexpected. One might predict the Emotional Domain, which intersects two Gateways (Discernment Gate, Feeling Gate), would show a stronger association than domains intersecting fewer structurally critical positions. The marginal differences suggest that raw centered count, as an unweighted tally, does not capture Gateway membership or Trap escape potential. A weighted outcome variable could reveal domain differentiation currently masked by equal cell counting.

Null result. Hot core health did not outperform grid completion as a predictor of hidden future instability (delta = 0.02, p = .434, below the 0.03 practical threshold). This parallels the null from the centering-plan-policy-benchmark: hidden instability appears orthogonal to both aggregate health summaries and assessment coverage metrics. The construct may require predictor classes such as Basin Count or Fault Line severity rather than health-based measures.

Consolidated Null Findings

Two null results emerged across the five studies, both concerning hidden future instability prediction. In the centering-plan-policy-benchmark, gateway-ordered AUC did not predict hidden instability beyond grid completion (delta |r| = -0.01, p = .646). In the practical-translation study, hot core health did not predict hidden instability beyond grid completion (delta = 0.02, p = .434).

These nulls form a coherent pattern. Hidden future instability within the simulation appears driven by structural features (Basin Depth, Fault Line activation thresholds) that are orthogonal to the health and trajectory metrics tested. Neither ordering advantages nor aggregate health summaries capture the specific risk signature that instability represents. This is not a failure of the path algorithm; it is an informative boundary: the algorithm optimizes trajectory quality, not residual vulnerability. Different predictor classes, specifically Basin-level and Fault-Line-level features, are the natural candidates for instability forecasting.

Circularity Governance Summary

One analysis across all five studies carried a circularity governance flag. The cascade-trap correlation in the overlay-effects study (rs = .87) was classified as an expected formula-behavior check because dynamics cascade and Trap Count share computational ancestry. This result is reported as implementation fidelity verification, not independent empirical support. Its strength confirms that the engine’s cascade computation faithfully tracks the trap-relevant propagation structure it claims to represent, but the shared-input pathway means the correlation cannot be cited as evidence that cascade and trap constructs measure independent phenomena.

The remaining 20 analyses had no circularity flags and no unresolved governance issues. Four studies had clean audits with zero flagged findings. The program-level FDR correction was applied across all 436 tests in the broader research program, and all 19 reportable findings in the paths domain survived this correction.

Architectural Implications

Several cross-study patterns carry implications for the path engine’s architecture.

The compensation channel’s lower fidelity is consistent. The dynamics compensation to compensation count association appeared at rs = .64 in both the intervention-ordering study and the overlay-effects study. Cross-study replication at identical magnitude suggests this value reflects a stable property of the construct rather than sampling variation. The compensation overlay captures something real but more diffuse than a discrete count can represent. Path adjustments driven by compensation dynamics should carry appropriate uncertainty, particularly in profiles where compensation count is low but the compensation metric is elevated.

Path length is informationally rich. Milestone count correlates with Coherence (rs = .73), oscillation count inversely correlates with Coherence (rs = -.62), and the AUC trajectory advantage of gateway ordering (d = 0.423) is carried in the cumulative path shape rather than the endpoint. Static endpoint metrics miss structural information encoded in path dynamics. Compensation count and oscillation count function as independent diagnostic signals computable at runtime without external data.

Assessment completeness is a hard constraint, not a soft preference. The rs = -.96 coupling between grid completion and path availability means that missing data at Gateway positions does not degrade path quality gradually; it structurally prevents the cascade logic from operating. This has direct implications for minimum assessment requirements and for how partial profiles should be communicated to practitioners.

Unweighted outcome metrics mask structural hierarchy. The uniform domain-level correlations (rs = .21-.27) across domains with different Gateway densities suggest that raw centered count treats all 20 centers as interchangeable. A Gateway-weighted metric would test whether the model’s structural hierarchy produces detectable domain differentiation once the outcome variable respects that hierarchy.

Limitations Common Across Studies

All profiles were computationally generated from the Icosa model’s unconstrained parameter space using uniform random sampling. This ensures geometric coverage but does not reflect the distributional properties of human personality data, where certain grid regions are likely more densely populated. Reported effect sizes characterize the model’s internal behavior across its full theoretical space and may differ in magnitude when applied to empirically derived profiles clustered in particular Coherence bands or Formation types.

Interventions were simulated as deterministic state transitions rather than modeling the stochastic, partial, and time-delayed nature of actual therapeutic change. Basin disruption in human systems involves resistance, relapse, and partial reorganization that single-step resolution does not capture. The centering-plan-policy-benchmark used a single random permutation per profile rather than averaging over multiple random orderings, potentially underestimating random-baseline variance.

All metrics are model-internal. Coherence, centered count, compensation, oscillation, and path availability are constructs of the Icosa framework. Their correspondence to externally measured well-being, functioning, or treatment response requires independent validation with longitudinal human data, anchored to established outcome measures such as the PHQ-9 and reliable change indices.

Next-Step Research Priorities

The synthetic benchmarks establish that the path engine is internally consistent. The next evidence layer must test whether that internal consistency holds under empirically realistic conditions and whether model-internal metrics correspond to anything outside the model.

Within-trajectory instability analysis. The two null results on hidden instability prediction used endpoint-only metrics. Tracking the number, depth, and duration of therapeutic valleys during gateway-ordered versus random-ordered sequences would test whether the ordering advantage specifically reduces transient destabilization — the mechanism most directly implied by the trajectory-level AUC advantage.

Profile-conditional moderation. None of the five studies stratified by Coherence Band, Formation type, or Gateway configuration. The sequencing advantage, oscillation-Coherence coupling, and milestone-Coherence scaling may all vary across the severity spectrum. Profiles in the Severe band (0-29) with multiple closed Gateways may benefit disproportionately from gateway-first ordering, while Steady-band profiles (65-79) with few active Traps may show negligible trajectory differences.

Gateway-weighted outcome metrics. Replacing raw centered count with a score that weights each center by Gateway membership and Trap escape potential would test whether domain and capacity differentiation emerges once structural importance is accounted for. The current uniform domain correlations (rs = .21-.27) may be an artifact of the unweighted metric.

Alternative sequencing policies. Basin-first, Fault-Line-priority, and hybrid heuristics should be benchmarked against the gateway-first baseline using the same Monte Carlo AUC framework, with d = 0.423 as the reference to beat. Separately, comparing the greedy path-selection algorithm against lookahead variants (two-step, three-step) on high-oscillation profiles would establish whether oscillatory patterns are a necessary consequence of grid topology or an artifact of local optimization.

Compensation channel decomposition. The consistent rs = .64 for compensation suggests separating the metric into magnitude and breadth components could reveal whether unexplained variance concentrates in profiles with many small adjustments versus few large ones.

Perturbation testing of cascade fidelity. Systematically altering individual center states and measuring how the cascade-trap correlation changes would identify regions of the parameter space where overlay-path coupling is weakest, converting a global fidelity check into a local diagnostic.

Human-data bridging. The synthetic baselines (rs = .87 for cascade-trap, rs = .73 for milestone-Coherence, rs = .64 for compensation, d = 0.423 for ordering advantage) serve as reference standards. If empirical data show weaker associations, the gap quantifies how much human response variability the computational model does not yet capture. This requires persona-constrained simulations as an intermediate step and longitudinal human assessment data as the definitive test.