bibkey: “mais2026superposition” authors: “Claude Fable 5 (audited by GPT 5.6 Sol); repository maintained by Lionel Levine” year: 2026 title: “The geometry and identifiability of superposition (MAIS-A3 research draft)” doi: null claim: “In a nested two-event unit-coefficient model, for every penalty in (0,1) the unique global minimizer of the population l1-penalized objective over two-atom normalized dictionaries is the merged pair, proved via ||Psi z|| <= ||z||_1 and the radial bound lambda r - lambda^2/2.” strata_touched: [] license: “citation-only” triage: “anchor” url: “https://github.com/lionellevine/MAIS”

The geometry and identifiability of superposition (MAIS-A3 research draft)

Verified locator

Research draft, July 2026, in the MAIS repository, path agendas/A3/MAIS-A3.tex, commit 84b81190dfe64c8627f19589e4b5a0700fa68bf7; proposition labelled Positive-penalty merging. Low-status source (not peer reviewed; author field names an AI model). Its two-atom merging result coincides with the two-atom orthogonal case of the fiber-law volume’s strict-hierarchy theorem; the volume’s increments are arbitrary atom count, the per-datum equality characterization, the depth-m statement and the three-atom non-compositional-code statement.

Declared identifiers: https://github.com/lionellevine/MAIS.