AMD+: A Design-Time Model-Selection Gate for Mixed-Intelligence Systems
Langiu (2026) introduced AMD, a specification language and methodology that types every service of a system by its intelligence mix and marks which services are served by a language model. We describe AMD+, the deterministic gate that turns such a specification into a model allocation before any code runs. For every model-served service the gate infers a capability-requirement vector from declared fields of the specification (interactions, contracts, fail modes, tools, counter-examples, declared payload size, provider tier), and matches it against a catalogue of candidate models whose cells carry a provenance mark. The cheapest model that meets the requirement on every axis is selected; when none does, the gate returns the closest model and the violated axes, which drives the variation loop of AMD. We give the inputs, the axes and their scale, every inference rule, the catalogue schema, the selection algorithm and a two-service example.
Authors
- Alessio Langiu (ORCID: https://orcid.org/0000-0002-5706-7500)
Institutions
- Istituto di Scienze Marine del Consiglio Nazionale delle Ricerche (IT)
- National Research Council (IT)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-02
- DOI
- https://doi.org/10.5281/zenodo.23885888
- Primary Topic
- Advanced Software Engineering Methodologies
- Type
- preprint