Preprint · v1.1.004 Sep 2026
Verification-Gated Local Model Cascading
A benchmark study of a fixed local FAST → BALANCED → BEST escalation policy for coding tasks, comparing adaptive routing against small-only and strong-only conditions.
Independent research · 2026
Systems that know when to verify.
Research into local model routing, modular intelligence, and evidence-aware systems. Ideas are published openly, with their methods, limits, and status made clear.
RESEARCH SYSTEM / 01
01 / ROUTESpecialist
02 / REVIEWVerifier
CONTROL PLANEEvidence
gate
gate
INPUTTask
OUTPUTClaim
Conceptual illustration · not a deployed system
01 / Selected work
Research releases
Preprints and concept work, presented with the scope and limits stated in each release.
Concept paper · v1.104 Aug 2026
Evolving Hierarchical Modular Intelligence Networks
A proposed architecture for specialist language models coordinated through evidence-carrying claims, separate verification gates, and hierarchical synthesis.
Research status. Both releases are preprints and have not been peer reviewed. The cascading results apply to the stated benchmark, model configuration, and operating point; they do not establish universal efficiency or predictive-routing advantages. The modular network paper is a proposal: it reports no implemented system or experimental results.
02 / How to read this work
Claims with boundaries.
Research should make its evidence legible, including what remains untested.
The current work follows one thread: route tasks to capable local models, then make verification and evidence handling explicit. The two releases represent distinct stages of that inquiry: a measured benchmark and an architectural proposal.
01
Measure a bounded question
Compare defined model conditions on a named benchmark and report the tested operating point.
02
Keep verification visible
Show where a verifier or test suite enters the workflow, and distinguish verification from model confidence.
03
Separate proposal from result
Architectural concepts are labeled as proposals until an implementation and experiments provide evidence.
Built in the open.
Research, software, and systems work from Mystrowin.
Generative AI assistance is disclosed in the Zenodo records. The author states that they directed the research question and design, reviewed the releases, and accept responsibility for their claims and conclusions. Paper text and original figures in the first record are licensed CC BY 4.0; software and third-party materials keep their own licenses. Refer to each Zenodo record for its applicable license and citation metadata.