BMT-08.03 Executive Summary#
BlueMirror.tech | May 2026#
James Okafor needed a cardiology referral. The Expert Exchange Layer had to answer a privacy question, not a routing question. What does the cardiologist need to know about James to help him? And what does the cardiologist have no business knowing?
The cardiologist needs the cardiac history, the full medication list, the allergy list, the recent vital signs, and the symptoms James reported conversationally. She does not need his financial situation, his family dynamics, his grocery preferences, his cognitive assessment scores, or his BGO Sage status. The system’s job is to draw the line.
The context packaging architecture begins with a computation called the Minimum Viable Context, which runs through a five-step process for every expert interaction. Query classification identifies the domain, urgency, and complexity. Expert matching retrieves the target expert’s capability schema with required and optional context fields. Consent intersection compares the expert’s requirements against the person’s active consent grants. Gap analysis identifies required fields blocked by missing consent and escalates to the person before proceeding. Package assembly extracts approved data from the MoC, formats it for the target expert’s interface (FHIR for clinicians, structured JSON for AI agents, readable summary for personal circle experts), and stages it for membrane transmission.
The minimum viable context is the intersection of what the expert needs and what the person has consented to share. If the expert requires data the person has not consented to share, the system escalates with a specific, plain-language request. The person decides. The system does not override.
Context packages come in three configurations determined by the person’s agency level. Minimal packaging provides only the query itself, used at FULL_HUMAN when the person has not authorized context sharing. Standard packaging provides the query plus required context from the capability schema, filtered by consent. Full packaging provides the query, required context, and optional context, used at TRUSTED_DELEGATION with broad consent. The three levels are computed per interaction based on the current agency level and consent state.
The architecture defines explicit exclusions that apply regardless of packaging level. The full Map of Context is never exposed to any external expert. The raw MoC contains cross-domain correlations, temporal patterns, and preference models that would reveal far more than any single expert needs. Medical conditions not relevant to the query are excluded. Financial information is excluded from all non-financial contexts unless the person consents to share it for a specific purpose. Cognitive assessment scores are excluded from all contexts except the clinician managing cognitive care, because those scores could be used against the person in legal, financial, or social settings.
The article walks through James’s cardiology referral as a worked example. The system assembles a package including cardiac history, medications, allergies, recent vital signs from wearables and home sensors, and conversationally reported symptoms. It excludes financial situation, family dynamics, grocery preferences, entertainment habits, cognitive scores, and BGO status. James reviews a plain-language summary on his phone and approves. The cardiologist receives a FHIR-formatted clinical summary containing 30 days of continuous vital signs that no referral packet has ever contained before.
The article names what context packaging cannot do. It cannot guarantee the expert uses the context appropriately after receiving it. It cannot perfectly determine relevance for novel queries that cross domain boundaries. The architecture handles these limitations through the audit trail and the person’s ability to review what was transmitted and adjust consent settings.
The full article is available at bluemirror.tech.