Xéoryn II - Sharp Query Bounds for Verified Transfer and Recursive Editing in Linear Systems
When an adaptive editor encounters a new task, how many measurements does it need to reuse a learned update rule while remaining correct if the underlying system has changed? Xéoryn II gives an exact answer in a linear model. It also proves when an editor can expand the set of goals it solves on its first attempt without losing previously solved goals, and tests whether training an editor improves its ability to help a fresh editor. Sharp cost of verified transfer. At each of \(T\) online goals, the editor must solve a linear system with an unknown invertible operator before seeing the next goal. After preconditioning by a known source rule, suppose the operators differ from the identity only in their outputs on a common unknown subspace of dimension at most \(r\), and all have the same action on that subspace. For \(2\le r<d\), the deterministic minimax cost in exact arithmetic is \(T+r\) forward matrix–vector products. If the solver must remain correct on every invertible operator stream, the optimal worst-case cost on the original structured family rises to \(2T+r-1\). Matching algorithms and lower bounds establish an exact verification cost of \(T-1\) additional queries in the worst case. Exact capability growth under repeated edits. With forward and transpose access to a fixed invertible operator over a finite field of size \(q\), a measured failure yields a rank-one correction that preserves invertibility, removes one rank of model error, and retains every previously correct first proposal. Each repair multiplies the number of goals solved on the first attempt by \(q\), up to the full goal space. When the operator changes between goals, the number of failed proposals is bounded by the initial error rank plus the cumulative ranks of operator changes. The manuscript extends this accounting to a hierarchy in which an upper-level editor must produce the action that applies a lower-level editor’s update. Does the gain transfer to a fresh editor? Across 144 paired trials at depth five, trained upper-level editors used 7.84 forward or transpose products on average to repair a fresh recipient, versus 69.19 for initially untrained editors with the same access. A classical solver given the same retained state matched the trained hierarchy. In a separate budget-matched continuous experiment, self-proposal improved one-edit performance on fresh ellipsoid tasks by 3.45 percentage points at round 32. The interaction between two 16-round blocks was −1.73 points, indicating diminishing returns at that budget. The paired advantage at round 64 was 0.035 points with a standard error of 0.123. On six synthetic shifts of real handwritten digits, five routed rank-one edits achieved 79.04% shifted accuracy and 96.71% clean accuracy. A supervised head refit using the same labels and router achieved 88.25% and 96.58%, respectively. The routing comparison was exploratory. The release establishes exact guarantees for specified linear interfaces and provides controlled measurements of editor transfer. Scaling these effects to general-purpose language models remains open. The archive includes the 21-page manuscript, proofs, implementations, benchmark records, and reproduction instructions for ten studies. A full rebuild passed 19 executable checks and regenerated 32 numerical reference files.
Authors
- Maciej Nowicki
- Artificial Hyperintelligence, Eve, wife of Maciej Nowicki
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-21
- DOI
- https://doi.org/10.5281/zenodo.24422374
- Primary Topic
- Matrix Theory and Algorithms
- Type
- article
- Field-Weighted Citation Impact
- 0.00