Evidence-Calibrated Hierarchical Repair of Combinational RTL under Bounded Verification Budgets

Verification-guided repair prioritizes candidate edits using evidence from failed executions, but its effectiveness under bounded verification budgets depends on both the fault model and the search schedule. We investigate evidence-calibrated hierarchical repair (ECHR), combining backward output-cone localization, four implemented semantic counterexample tags, operator-family priors, feedback-based reranking, guarded two-edit branches, and exact finite-domain acceptance. The tag classifier, operator priors and cone localization adapt the counterexample-signature machinery of the VGMA verification-guided multi-agent RTL framework. Across ten deterministic synthetic-corpus seeds, we execute 2,520 detectable faults on 180 small combinational reference circuits, including 1,080 two-site faults. Every candidate is evaluated on 64 witness vectors, and accepted repairs must agree with the reference for all 512 three-input, three-bit assignments. With a fixed 30-query budget, ECHR repairs 820/2,520 cases (32.54%), compared with 400/2,520 (15.87%) for random search, 762/2,520 (30.24%) for static combined ranking and 771/2,520 (30.60%) for the strongest static ranking (tag-only). The ECHR–combined-ranking difference is positive in all ten seeded replications (mean +2.30 percentage points; 95% seed-bootstrap interval +1.67 to +2.94). Ablations show complementary effects of feedback reranking and two-edit expansion, although two-site repair remains limited to 53/1,080 cases. Results characterize a restricted, synthetic bit-vector repair problem rather than sequential or industrial Verilog.