Metamorphic Coverage and Reward Sensitivity in Agent Programs
Year: 2026
Type: preprint
Abstract: We test whether a repaired verifier accepts archived agent programs that fail fresh, reference-qualified input probes. The study covers 100 programs for one session-window component. Before execution, we freeze 120 traces combining six source-name bijections, five positive event-time scales and four offsets. A shipped reference and two planted controls qualify the harness. Each case runs in a fresh isolated process. All 12,000 program–case executions complete. Ninety-eight programs match every declared contract vector; two differ at identity and every transformed case. All 100 satisfy own-anchor equivariance. Source inspection explains the two differing vectors through alternative event-count idleness policies, which the short task design does not fully resolve. The 119 nonidentity cases add no departures among identity-conforming programs. Replication of baseline-to-repaired scoring changes two labels and ten standardized reward coefficients on fixed evaluation cohorts; probe augmentation changes none. A census of 794 public records supplies archival context. The measurements concern saved programs and include no policy update. We propose an unrun matched private-task experiment to test whether verifier repair improves training outcomes at a defensible cost. Materials. The PDF, editable source, vector figures and sanitized observations accompany the reproduction code. Public benchmark material remains evaluation-only. This is a preprint and has not been peer reviewed. Computational assistance. OpenAI Codex, including GPT-6.1-Sol, assisted study design, literature review, implementation, analysis, figure generation and manuscript preparation. Reported measurements come from logged code executions; the record includes the code and observations needed to reproduce the analysis.
Language: English
Subfield: Software
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225578214