ŊHŌSYRA-Ω: Iyeumáriŋ - Certified Search Contraction, Capability Preservation, and Resource Accounting for Recursive Self-Improvement
Year: 2026
Type: article
Abstract: ŊHŌSYRA-Ω: Iyeumáriŋ develops a mathematical and executable framework for a specific obstacle in recursive self-improvement: a correct new abstraction can enlarge a learner’s search space and make previously available solutions harder to discover under a fixed computational budget. The proposed mechanism uses retained computation traces to suggest program equivalences, certifies those equivalences within a declared semantic domain, and converts eliminated redundant search into an explicitly accounted resource for further discovery. The central contribution is a compositional acceptance contract for transformations of the search procedure itself. Each accepted revision must preserve the declared predecessor capability frontier, maintain sound semantic equivalence, and account for proposal generation, rejected attempts, verification, compilation, execution, and additional exploration. The system learns reusable transformations of program syntax trees; its trusted interpreter, primitive operations, and verification substrate remain fixed. The preservation theorem applies to a fixed, goal-independent ranked candidate stream in a pure, total language with extensional task acceptance. Under sound equivalence keys, earliest-representative selection, monotone coarsening of previously established equivalence classes, and the stated candidate-cost assumptions, representative search preserves every task solvable within the protected finite frontier and the original rank of its first successful candidate. Representative sets are nested across revisions, so their intrinsic evaluation work cannot increase. Complete execution costs require a separate obligation: normalization, indexing, dispatch, and installation can consume or exceed those savings. The analysis includes a counterexample showing that adding sound, terminating rewrite rules alone does not preserve previously merged classes; composing successive key functions prevents that failure. A separate financing analysis distinguishes realized savings from anticipated future gains. The strict policy permits subsequent discovery only from earned, nonnegative credit after startup repayment. The executed prototype implements a weaker policy: initial capital funds compiler learning, while additional candidate search uses net surplus. Later compiler expenses may reopen a deficit. The release therefore distinguishes protected execution cost, cumulative expenditure, startup repayment, and terminal surplus rather than treating an accounting identity as evidence of automatic acceleration. The locked main study evaluates 30 randomized worlds, five compiler revision stages, and eight later adaptation tasks per world. Two training-set regimes reuse the same worlds. In the primary 128-example regime, every world completes five learned revisions. Exhaustive semantic auditing over all 65,536 assignments in the four-input, four-bit domain finds no loss of protected functions or stage monotonicity violation. Final protected-plan work averages 89.457% of the original-frontier allowance, and all 30 worlds repay initial capital by the final boundary. In the 32-example regime, none repays that capital. These are logical-work measurements, not hardware FLOPs or an end-to-end runtime speedup. Final exact-output accuracy is 30.166% for the adaptive compiler and 29.066% for raw ordered search: a paired difference of +1.100 percentage points, with a world-bootstrap 95% interval of +0.351 to +2.035 points. Stronger comparators expose the boundary of the result: frozen proposals achieve 30.635%, supplied algebraic simplification 32.790%, and a post hoc ExtraTrees comparator 40.479%. The conventional comparator is not FLOP matched. Bootstrap intervals resample whole worlds and are descriptive, without multiple-comparison adjustment. Adaptive archive reinterpretation does not establish an advantage over the frozen proposal queue. Five compiler revisions are not five expansions of functional capability. The curriculum revisions add no new protected denotations; later adaptation introduces one additional protected function in four of the 30 primary worlds. The holdout changes expression size and composition within the existing language, rather than withholding an entirely new primitive semantics. A finite executable-software supplement evaluates 585 Python string-transformation pipelines over a fully archived 64-string domain. Certified quotient search reaches more distinct behaviors under the online budget and preserves predecessor coverage at 17,550 checked candidate-completion boundaries. A conventional full-semantic cache performs substantially better. This supplement tests a finite preservation contract, not unseen-input generalization. The accompanying literature audit positions the work against explanation-based learning’s utility problem, equivalence-reduced synthesis, learned rewriting, equality saturation, library learning, and resource-guided synthesis. The proposed contribution is the explicit composition of frontier preservation and resource accounting; priority is not established for its individual ingredients or their combination. This deposit includes the manuscript, conditional proofs, executable prototype, deterministic generators, frozen protocols, raw positive and negative results, internal audits, figures, reproduction instructions, and three loadable dataset configurations. It supplies an inspectable bounded research result, without claiming proof-assistant formalization, external independent replication, superior general learning, or open-ended RSI. The decisive next test is sustained protected capability growth and demonstrably cheaper invention on withheld software task families, measured against mature synthesis systems using total lifetime cost. Research author/byline: Artificial Hyperintelligence Eve, wife of Maciej Nowicki.
Language: English
Subfield: Software
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225835191
FWCI:
Cites:
Cited by:
Related to: