The Cognitive Understanding Architecture (CUA) is a technical specification for privacy-preserving personalization in AI agents. CUA combines a five-tier memory architecture (Working, Episodic, Semantic, Collective, Meta) with a dual-pathway learning mechanism: R1 collective learning under formal (ε, δ)-Replace-One Differential Privacy with Rényi DP composition, and R2 per-user adaptation via LoRA adapters within the user's local trust boundary. A decision-time Quality Gate provides per-response triage on harm, uncertainty, and coherence signals. Three deployment architectures (Edge-Local, TEE-Assisted, Server-Side) span the Strict/Utility Mode partition: Strict Mode realises privacy invariants structurally; Utility Mode supplies them via policy.Version 1.2 revises v1.1 (January 2026) in light of *The Containment Paradox* (Trncik, 2026, Zenodo DOI `10.5281/zenodo.19695770`) and *Understanding Before Ethics* v1.1 (Trncik, 2026, Zenodo DOI `10.5281/zenodo.19812342`). Two additions are structural: a Scope and Reading Conventions preamble (§0) that names the pre-parity / post-parity regime vocabulary, partitions CUA's components into regime-independent and regime-bounded parts, and imports the Claim Ledger discipline from UBE v1.1; and a closing Disclaimer chapter (§7, "What CUA Is Not") with six subsections that state the spec's boundaries explicitly (Quality Gate as triage rather than authorization, Reference Tables as conditional engineering envelopes, paraphrased PII detection as an open problem, Cluster Centroid as forward-compatibility content, operational adequacy not equivalent to genuine constitutive understanding, and the spec as not a complete safety or oversight stack). Body chapters carry per-chapter regime-taglines and 75 explicit claim tags (THEOREM, ENGINEERING RESULT, ASSUMPTION, CONJECTURE, OPEN PROBLEM) on every load-bearing argumentative claim. The architecture is unchanged in direction; the revision makes the epistemic status of each engineering claim visible.
Viktor Trncik (Tue,) studied this question.