The Same Label, Different Relationships: What 35,817 Personality Pairings Show About Dyadic Prediction

Two Suspended People in a Relationship

Two people both classified as Suspended — a formation characterized by holding patterns and unresolved commitment — pair up. What kind of relationship do they produce?

If personality labels were predictive, the answer would be one thing. But Icosa Atlas computed every possible Suspended+Suspended pairing across its formation recipe library, and the answer is 12 different things: Mutual Pursuit, Fused, Clashing, Working Through, Shifting, Disorganized, Productive Tension, Danger, Shared Delusion, Power Gap, Enmeshed, and Lifting. Twelve structurally distinct relational configurations from the same two-word label on both sides.

Enumerating 35,817 pairings across 83 formation recipes, 165 clinical personas, and 8 crosswalk systems (MBTI, Big Five, Enneagram, DISC, Holland, Type A/B/C/D, Temperaments, and Attachment) shows this pattern everywhere. 86% of same-structural-formation pairs produce multiple different dyadic formations depending on the capacity-level details that the structural label collapses away.

And 35,817 is a small number. Each person’s profile is continuous — shaped by their specific responses, not snapped to a preset type. The real combination space across any population is effectively infinite. This enumeration covers the structured library; actual human pairings would produce even more variation. Two people both labeled Engaged can end up in Working Through or a Shared Wound. Two Balanced+Narrowed pairings can produce five different relational outcomes, from Mutual Pursuit to Lifting. The structural label is a summary, and summaries lose information — and the information that’s lost turns out to be what determines relational outcomes.

How Much Does the Relationship Add?

If individual profiles don’t fully predict what the relationship looks like, how much is left over? Icosa’s semantic embedding system gives a direct measurement.

The system maps both individual profiles and dyadic profiles into a shared 3,072-dimensional vector space. For each pair, the analysis computed a midpoint of the two individual embeddings (the “summed inputs” prediction) and compared it against the direct dyadic embedding (what the engine actually produces when both profiles interact). The cosine similarity between those two vectors measures how much of the relationship is predictable from the individuals alone.

Across 4,923 pairs tested, the mean cosine similarity was 0.73 with a standard deviation of just 0.027. The range ran from 0.64 to 0.84.

Three-quarters of the relationship’s structural direction is inferable from knowing both individuals. The other quarter isn’t noise. It’s the emergent relational architecture: trap interactions that only exist between two people, cross-compensations that activate when one partner’s strength meets the other’s vulnerability, gateway configurations that open or close depending on both profiles simultaneously. None of that exists in either individual profile, and averaging two profiles together won’t surface it.

The consistency across dyadic families is worth noting. Resonant relationships (mean 0.74), Distressed relationships (0.73), and Rupture relationships (0.73) are all predicted from their inputs at roughly the same rate. Healthy and unhealthy relationships are equally irreducible to their components.

The Attractor Problem

The 3,486 formation-recipe pairs don’t distribute evenly across the 45 dyadic formations. They cluster.

Working Through absorbs 715 pairs — just over 20% of all formation combinations. Lifting takes 510. Mutual Pursuit takes 310. Between them, three formations account for 44% of all outcomes. At the other end, Parallel Lives appears exactly once. Shared Blindness appears twice. Cemented appears twice. Controlling appears three times.

Most individual formation recipes produce profiles in the moderate-coherence range. When two moderate profiles pair, the result is typically a moderate relational outcome: Working Through (functional but effortful), Lifting (one partner supporting the other), Mutual Pursuit (enmeshed but active). The extreme dyadic formations require extreme individual inputs, and those are rare in the library.

Clinicians using Icosa will see Working Through constantly and Parallel Lives almost never. That recalibrates what’s informative. A Working Through classification is the relational equivalent of room temperature. A Parallel Lives or Controlling classification signals a structurally unusual configuration worth close attention.

Some People Have More Relational Range

Not all individual formations produce the same diversity of relational outcomes.

Suspended produces 24 different dyadic formations across its 924 pairings. Pair a Suspended individual with enough different partners and nearly the entire dyadic space opens up. Holding produces 23 outcomes. Narrowed and Poised each produce 22.

At the other end, Dispersed produces only 6 from 84 pairings. Flaring produces 10. These formations are structurally extreme enough to dominate the dyadic interaction regardless of what the partner brings. A Dispersed individual’s relational range is narrow because the profile’s severity collapses most partner combinations into a small set of distressed configurations.

Partner selection matters more for some formations than others. A Suspended individual’s relational prognosis is highly partner-dependent. A Dispersed individual’s is driven primarily by the severity of the individual profile, and stabilization there has to come before partner dynamics become clinically meaningful.

Composite Signatures by Family

The 10 dyadic composite metrics — repair capacity, dyadic vitality, healing readiness, pain-pathology ratio, stress vulnerability, mutual perception, differentiation, cross-compensation dependency, shared stability, and therapeutic fit — produce distinct signatures across the eight dyadic families. Two stand out as clean discriminators.

Resonant relationships: repair capacity 0.55, stress vulnerability 0.37. The highest repair and lowest stress fragility of any family. These couples can absorb conflict and metabolize it.

Rupture relationships: repair capacity 0.32, stress vulnerability 0.63. The inverse. Low recovery capacity, high fragility.

Transitional relationships sit at the midpoint: 0.41 and 0.41. In motion, but the direction isn’t set.

Distressed and Mirrored both show elevated stress vulnerability (0.58) but through different mechanisms. Distressed couples have low repair (0.34) and the relationship is deteriorating. Mirrored couples have slightly higher repair (0.38), but their mirrored structure means both partners’ vulnerabilities fire at the same time, with no one left to stabilize the system.

Asymmetric couples show moderate repair (0.44) but uneven stress vulnerability (0.46) — one partner typically carries more of the load. Complementary couples: repair at 0.39, stress vulnerability at 0.53. The complementarity helps with daily functioning but doesn’t translate into resilience under pressure. Stagnant couples: repair at 0.36, stress vulnerability at 0.51 — flat, stable, not growing. The composites capture the difference between “not in crisis” and “not going anywhere.”

A couple with repair capacity above 0.50 can likely tolerate the discomfort of active therapeutic work. Below 0.35, individual stabilization may need to come first.

The 73% Boundary

The embedding compositionality result (cosine 0.73, σ = 0.027) sets a clean boundary on what individual assessment can and can’t do for relationship prediction.

It can get you into the right neighborhood. Knowing both partners’ full profiles predicts roughly three-quarters of the relationship’s structural signature — enough to make informed guesses about dyadic family and approximate composite ranges.

It can’t get you the rest. That other 27% requires actually computing the dyadic interaction: running both profiles through the engine, computing cross-partner trap interactions, gateway configurations, and tensor alignments that only come into existence when two specific profiles meet. The clinically important constructs — the specific trap interactions, the gateway blockages, the repair pathways — live disproportionately in that emergent quarter.

Individual profiles and dyadic profiles measure different things. Knowing both people well gets you most of the way there, but the relational structure that emerges when two specific profiles interact can’t be reconstructed from the parts.