Answers are grounded in this paper and cite their sources.
How can a self-modifying system certify that a rewrite improves its objective when the rewrite introduces outcomes its evaluator has never scored? Iryŏx II studies this problem under corrupted pairwise comparisons. New outcomes may have independent unknown utilities outside the old ontology’s convex hull. Each new group is grounded through a cycle of comparisons sharing one persistent operational reference. The main theorem constructs an online evaluator that keeps each score fixed once published and bounds the absolute error of every pairwise score difference by \(2τ\) across arbitrarily many expansions. Per cycle, at most one initial comparison can be arbitrarily corrupted and the remaining initial comparisons are exact. Honest local follow-ups have absolute error at most \(τ\), with up to \(b\) arbitrary replies per cycle. The evaluator uses at most \((2b+1)log_2 m\) follow-ups for a cycle of \(m\) outcomes. Two cycles already force the \(2τ\) bound in the worst case, even with unlimited local follow-ups. A reported improvement greater than \(2τ\) therefore certifies a true improvement on the operational objective. The manuscript also determines the value of a partial audit. For a cycle with exact healthy initial readings, signed discrepancy \(S\), and \(h\) fault locations compatible with the audit transcript, the exact minimax error of one coherent score vector across all pairs is \( S(1-1/h)\). This quantifies what the evaluator can certify before identifying the faulty comparison. A linear-programming formulation extends the analysis to finite feature models with grouped measurement faults and identifies when incomplete observations leave the worst-case error for a new score difference unbounded. A reproducible experiment tests the evaluator on six modular exponentiation tasks. Across 33 proposal rounds, each with one corrupted initial comparison and one false follow-up, the evaluator accepts 32 strictly improving rewrites. The final plans use 49 modular multiplications, down from 430 for the initial plans and 15.5% below the 58 required by binary exponentiation. The proposal mechanism is fixed, the faults are simulated, and the reported gains are in operation count. The release includes the full proofs, executable benchmark, round-by-round trace, exact rational-arithmetic checkers, 19 finite symbolic SMT checks, and a linear-programming cross-check. For recursive self-improvement research, the result gives an explicit condition under which an evaluator can certify successive improvements as the range of outcomes it must score expands.
No takes yet. Share an insight, caveat, or question.
Nowicki et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: