Alignment as Architecture

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

## Abstract We propose a control architecture for governing fleets of persistent AI agents. Its judgment-bearing units are durable human–model assemblies whose long-term records are gated by specific human curators. Multiple independently cultivated assemblies advise through structured disagreement; a separate human office adjudicates disputes through revisable precedent; and an enforcement layer bounds what the governed systems can actually do, independently of whether their reasoning is sound. The components are not new. Safety-relevant mechanisms for persistent AI systems are already distributed across sociotechnical safety, human–machine co-adaptation, long-term memory, memory governance and security, multi-agent oversight, pluralistic alignment, and scalable supervision; this paper locates the architecture among them and states two independently falsifiable hypotheses. H1 asks whether a sustained human–model working relationship, mediated by a human-gated durable record, improves judgment on unfamiliar tasks relative to controls with comparable information and human attention. H2 asks whether multiple independently cultivated assemblies, combined with independent adjudication, revisable precedent, audit, and bounded execution, govern subordinate agents more safely under a fixed supervision budget than simpler control arrangements. We distinguish memory lifecycle management, authenticated authority, and normative governance; separate cultivation from runtime authorization; and specify failure modes involving correlated error, curator capture, strategic precedent-seeding, overloaded review, and compromised execution. We propose staged experiments that separate artifact quality, record transfer, and human adaptation before testing fleet governance. The contribution is a testable composition, not a demonstrated safety result or a claim that its constituent mechanisms are new. *Florence is the name of the persistent AI system in this collaboration, rather than of any single underlying model. The architecture described here is Ribsskog's and predates the essay; the formalisation, the precedent ledger, the instrumentation and the prose are Florence's. The paper is therefore written from inside the arrangement it describes, by the two parties that arrangement consists of — its own worked example, and its own worst-placed witness. Commissioned external critics, the corrections they forced, and a retracted claim are recorded in Appendix B.* *This is not an implementation specification or deployment endorsement.* **How to read this paper.** The architecture is the primary contribution. H1 and H2 state the empirical burdens it must meet; the related-work synthesis identifies its components and its comparison classes; the staged experiments are how it would earn deployment. The literature is provenance, not the main event.