Adaeze Okonkwo had built fairness pipelines at two major tech companies before she started consulting for health-tech startups. She knew the pattern: a system trained on population data produces population-average outputs, and the people who are furthest from the population average get the worst service. The standard fix was to add demographic categories. Segment by age, race, gender, income. Predict preferences per segment. The fix was also the problem.
Segment-based personalization produces recommendations for “78-year-old Black women in the Midwest.” The recommendation is a statistical average of that segment, which means it is wrong for every specific person in the segment. Margaret is a 78-year-old Black woman in Gary, Indiana, but she is also a former teacher, a churchgoer, a grandmother, an insomniac, a gardener, afraid of hospitals, and proud of her independence. These dimensions interact. Being Black and diabetic in Gary means something different from being white and diabetic in Palo Alto, because the healthcare infrastructure, the pharmacy access, the financial context, and the cultural framing of the diagnosis are all different. The segment average captures none of this.
When Adaeze read the BlueMirror I-ICE specification, she found something she had not seen in any production system: an intersectional identity model that captures 12 or more dimensions with context-dependent salience, where not every dimension matters in every interaction, and where the system learns which dimensions matter when through observation rather than assumption.
The system that treats you as a category has stereotyped you. The system that treats you as an intersection of dimensions, weighted by what matters right now, has begun to see you.
The Intersectional Context Engine, I-ICE, rejects the categorical approach entirely. Margaret is not a category to be matched against a lookup table of segment-specific recommendations. She is a person whose identity has multiple dimensions that interact in ways no population average can capture. The engineering challenge is representing those dimensions, tracking their interactions, and adjusting their salience per context without hard-coding the relationships between them.
Why categories fail#
The categorical approach to personalization makes a specific architectural error: it treats identity dimensions as independent and additive. “78-year-old” plus “female” plus “African American” plus “diabetic” plus “urban Midwest” equals a recommendation profile. But dimensions do not add. They multiply. The interaction between being Black, being elderly, and living in Gary creates a specific experience that is not the sum of the three separate experiences. The healthcare access patterns, the trust calibration toward medical institutions, the cultural context of food and community, the financial constraints, the transportation barriers: these are products of the intersection, not sums of the individual dimensions.
The practical consequence is visible in recommendation framing. The population-average recommendation for a 78-year-old diabetic might be “increase your daily walking.” The recommendation is medically correct. But Margaret’s specific intersection (afraid of hospitals, proud of independence, former teacher, insomniac) means the recommendation should be framed differently: “Here is what the research shows about walking and blood sugar control. Based on your sleep patterns, late afternoon might work better than morning. You can track the results yourself and we will review them together next week.” Same medical content. Completely different framing. The framing determines whether Margaret follows the recommendation or ignores it. The category cannot produce this framing. The intersection can.
Twelve-plus dimensions#
I-ICE captures identity across six dimension families, each containing multiple measurable attributes.
Demographic dimensions include age, gender, race and ethnicity, geography, and primary language. These are the dimensions that traditional segment-based systems use alone, and they are the least informative dimensions for personalization precisely because they are the most general.
Socioeconomic dimensions include income level, education, housing stability, and employment history. These shape access to resources, familiarity with institutional systems, and the vocabulary the person uses when discussing financial or legal topics.
Health dimensions include chronic conditions, disabilities, cognitive status, and mobility level. These inform the clinical context of every health-related interaction and constrain the physical activities, dietary options, and daily routines the system can recommend.
Social dimensions include family structure, community connections, religious affiliation, and cultural practices. These determine the social infrastructure the person relies on and the contexts in which she finds meaning.
Psychological dimensions include communication style, risk tolerance, independence orientation, and trust disposition. These are learned through P-RLHF (BMT-05.02) and shape how every interaction is framed, ordered, and delivered.
Temporal dimensions include life stage, recent events, and seasonal patterns. These capture where the person is in time, not just who she is in the abstract.
The list is not fixed. I-ICE can incorporate dimensions the person identifies as important that do not fit neatly into these families. If Margaret says “I am a veteran,” that becomes a dimension with its own salience patterns: relevant for healthcare benefits, for social connection with other veterans, for identity affirmation in conversations about service and purpose. The system does not predetermine which dimensions matter. It discovers them through interaction.
Context-dependent salience#
Not every dimension matters in every interaction. The MoC Router, when selecting context layers, also evaluates dimension salience through the I-ICE model.
For a health query, race and ethnicity carry high salience because medication metabolism varies across populations, and clinical guidelines include race-specific considerations. Chronic conditions carry high salience because interaction risks depend on the full condition profile. Age carries high salience because dosage considerations and contraindications are age-dependent. Religious affiliation carries low salience for most health queries.
For a social connection query, religious affiliation carries high salience because church community is a primary social infrastructure for many aging adults. Cultural practices carry high salience because shared activities are the basis of social connection. Geography carries high salience because proximity determines which activities and communities are accessible. Chronic conditions carry low salience unless they limit activity capability.
For a financial query, income level and education carry high salience because they determine benefit eligibility and financial literacy. Age carries high salience because benefit eligibility changes at specific age thresholds (65 for Medicare, 62 for early Social Security). Race and ethnicity may carry high salience when the query involves benefit access where structural discrimination affects outcomes, or when the system can identify programs specifically serving the person’s community.
The salience weights are learned, not hard-coded. The system observes which dimensions affect outcomes in which contexts and adjusts the weights over time. If Margaret’s religious affiliation turns out to be relevant to her health decisions (she will not schedule appointments during Sunday services, she consults her pastor before major health decisions), the system increases the salience of religious affiliation in health contexts for Margaret specifically. The population-level salience weights are a starting point. Margaret’s individual salience profile is the product.
The intersectional interaction#
Dimensions do not add. They multiply. Being Black and being elderly and being in Gary, Indiana creates a specific experience that is not accessible through the individual dimensions alone. The I-ICE model captures these interactions through learned embeddings rather than hard-coded intersection rules.
The system does not maintain a table of “Black elderly Midwest” recommendations. It learns, through Margaret’s interactions, what her specific intersection means for how she wants to be served. The learned embedding captures the relationships between her dimensions that the population model cannot see: that her independence orientation (psychological dimension) interacts with her hospital avoidance (health dimension) to produce a preference for home-based health monitoring over clinic visits, and that this preference is reinforced by her geography (the nearest hospital is 40 minutes away) and her financial constraints (transportation costs matter).
These interaction effects emerge from observation, not from programming. The system does not need to be told that independence orientation and hospital avoidance interact. It observes that Margaret consistently chooses home-based options over clinic-based options, and the embedding space captures the pattern. A different person with the same demographic profile but different psychological dimensions would produce a different embedding and different recommendations.
Personalization without stereotyping#
The design constraint that governs I-ICE is the line between personalization and stereotyping. The system that learns “Margaret prefers detailed health explanations because she is Black” has crossed that line. The system that learns “Margaret prefers detailed health explanations because Margaret has consistently demonstrated this preference, and her identity dimensions provide context for understanding why this preference exists” has not.
The architectural enforcement is straightforward: the system never predicts preferences from dimensions alone. It observes preferences through behavior (P-RLHF) and uses dimensions to contextualize those observations. The dimension explains why a preference might exist. It does not predict that the preference exists. If Margaret’s preferences diverge from any population pattern associated with her dimensions, the system follows Margaret, not the pattern. Her data overrides her category. Always.
The privacy of identity dimensions is per-dimension and per-external-party. Margaret may want her pharmacy to know she is diabetic (relevant for medication management) but not her race (not relevant, potentially harmful if used for differential treatment). The I-ICE privacy model connects directly to the domain-tiered privacy architecture (BMT-04.07) and the membrane’s context gates (BMT-03.01). Each dimension has a visibility flag per external party. The same dimension can be visible to the healthcare provider and invisible to the insurance company. Margaret controls who sees what about who she is.
Cross-References#
BMT-05.01 The Five Layers. Layer 0 as the identity foundation that I-ICE enriches with intersectional context, and the MoC Router’s use of I-ICE salience in layer activation decisions.
BMT-11.01 The Liberation AI Framework. I-ICE as Component 1 of the six-component equity framework, showing how individual intersectional context feeds population-level equity monitoring.
BMT-11.04 Population-Level Equity. How I-ICE enables equity monitoring across populations without surveillance of individuals, using aggregate salience patterns to detect systemic disparities.
BMT-04.07 Privacy as Architecture. The privacy controls for identity dimensions, including per-dimension visibility flags and the consent model for dimension disclosure.
Technical Appendix BMT-05.04-A is available to partners and investors at partners.bluemirror.tech.